# Can My Agent Use corpus The first review covers selected current-version claims. Every other cell stays unknown until a public source is reviewed. ## Accountable human handoff HTML: https://canmyagentuse.com/features/human-handoff Markdown: https://canmyagentuse.com/features/human-handoff.md This row covers an accountable transfer of responsibility, not merely alerting someone that a run exists. A useful handoff identifies the recipient or queue, current status, requested decision or action, relevant context and artifacts, deadline, permissions, and who owns the next step. Evidence should record assignment and reassignment, acknowledgement, comments, escalation, service-level timers, notification channels, redaction, audit events, resume semantics after a decision, and whether the agent can continue autonomously while the handoff is pending. --- ## Administrative policy controls HTML: https://canmyagentuse.com/features/admin-policy-controls Markdown: https://canmyagentuse.com/features/admin-policy-controls.md This row covers centrally enforced organization policy, not a user preference or a natural-language instruction the agent may ignore. Useful controls cover allowed models and providers, tools and connectors, MCP servers, network destinations, data sharing, uploads, memory, retention, training use, public links, autonomous actions, and sub-agent fan-out. Evidence should record scope, role required to change policy, inheritance and exceptions, client support, propagation time, offline behavior, precedence over local configuration, change logs, and the user experience when policy blocks an action. --- ## Agent Plugins failure isolation HTML: https://canmyagentuse.com/features/agent-plugin-failure-isolation Markdown: https://canmyagentuse.com/features/agent-plugin-failure-isolation.md Reject invalid components narrowly without disabling valid components in the same package. This row remains unknown until exact, dated evidence covers this behavior. --- ## Agent Plugins packaged MCP HTML: https://canmyagentuse.com/features/agent-plugin-mcp Markdown: https://canmyagentuse.com/features/agent-plugin-mcp.md Load MCP server configuration contained in an Agent Plugins package. The Grok Bot cell is based on the Agent Plugins project's dated compatible-client registry. Individual MCP primitives, authentication flows, and extension fields remain separate claims. --- ## Agent Plugins packaged skills HTML: https://canmyagentuse.com/features/agent-plugin-skills Markdown: https://canmyagentuse.com/features/agent-plugin-skills.md Load portable Agent Skills contained in an Agent Plugins package. The Grok Bot cell is based on the Agent Plugins project's dated compatible-client registry. Scripts, references, progressive loading, and installation scopes remain separate Agent Skills claims. --- ## Agent Plugins portable manifest HTML: https://canmyagentuse.com/features/agent-plugin-core Markdown: https://canmyagentuse.com/features/agent-plugin-core.md Load the published portable plugin.json manifest and its core components. The Grok Bot cell is based on the Agent Plugins project's dated compatible-client registry. Unreviewed clients remain unknown, and the listing does not prove every optional client extension. --- ## Agent Skills HTML: https://canmyagentuse.com/features/agent-skills Markdown: https://canmyagentuse.com/features/agent-skills.md Recognizing `SKILL.md` does not prove progressive disclosure, script execution, supporting-file access, manual invocation, or identical project and user scopes. These are tracked as separate child capabilities. --- ## Agent Skills core and discovery HTML: https://canmyagentuse.com/features/skills Markdown: https://canmyagentuse.com/features/skills.md Agent Skills are reusable directories centered on `SKILL.md`, with optional scripts, references, and assets. The catalog treats basic discovery as narrower than full support for progressive loading, script execution, every install scope, or every extension field. Sourced cells therefore retain beta, plan, policy, code-execution, and surface conditions. A skill package can contain executable code; documented loading does not imply that Can My Agent Use executed or security-reviewed third-party skills. --- ## Agent Skills manual invocation HTML: https://canmyagentuse.com/features/skill-manual-invocation Markdown: https://canmyagentuse.com/features/skill-manual-invocation.md Let a user or agent explicitly select and activate a named skill. This row remains unknown until exact, dated evidence covers this behavior. --- ## Agent Skills progressive loading HTML: https://canmyagentuse.com/features/skill-progressive-loading Markdown: https://canmyagentuse.com/features/skill-progressive-loading.md Load skill metadata first and defer full instructions or resources until relevant. This row remains unknown until exact, dated evidence covers this behavior. --- ## Agent Skills references and assets HTML: https://canmyagentuse.com/features/skill-references-assets Markdown: https://canmyagentuse.com/features/skill-references-assets.md Resolve supporting reference files, templates, and assets from a skill package. This row remains unknown until exact, dated evidence covers this behavior. --- ## Agent Skills scripts HTML: https://canmyagentuse.com/features/skill-scripts Markdown: https://canmyagentuse.com/features/skill-scripts.md Resolve and execute optional scripts included with an Agent Skill under explicit runtime policy. This row remains unknown until exact, dated evidence covers this behavior. --- ## AGENTS.md root instructions HTML: https://canmyagentuse.com/features/agents-md Markdown: https://canmyagentuse.com/features/agents-md.md Discover and apply the cross-agent AGENTS.md convention at repository scope. This row remains unknown until exact, dated evidence covers this behavior. --- ## Artifact and workspace export HTML: https://canmyagentuse.com/features/artifact-export Markdown: https://canmyagentuse.com/features/artifact-export.md This row covers outputs that exist beyond rendered chat: files, patches, commits, repositories, notebooks, documents, slides, spreadsheets, archives, images, audio, video, and structured result bundles. A screenshot or copied code block is only partial when the original artifact has richer structure. Evidence should record download and API paths, git or cloud-storage transfer, archive structure, original versus converted formats, metadata and citations, checksums, size limits, expiry, access control, provenance from sub-agents, and whether an export remains usable without the originating harness. --- ## Audio file upload and understanding HTML: https://canmyagentuse.com/features/audio-file-input Markdown: https://canmyagentuse.com/features/audio-file-input.md This row covers recorded audio supplied as a file, not a live duplex voice session. Evidence should distinguish speech transcription from speaker diarization, timestamps, language identification, non-speech sound analysis, emotion or prosody claims, and native audio-model input. Record formats, channels, maximum bytes and duration, language limits, transcript editability, retention, model and plan requirements, and whether tools or sub-agents receive the original recording or only a derivative transcript. --- ## Audit logs and actor attribution HTML: https://canmyagentuse.com/features/audit-logs Markdown: https://canmyagentuse.com/features/audit-logs.md This row asks whether security-relevant activity is recorded with enough identity and provenance for investigation. Conversation history is not sufficient when it omits tool parameters, connector reads, file changes, approvals, child-agent actions, model routing, policy decisions, sharing, exports, and administrator changes. Evidence should record event types and fields, user and service identities, timestamps, model and harness version, retention, search, export or streaming API, SIEM integration, tenant isolation, tamper resistance, redaction, regional placement, and documented gaps. --- ## Automatic context compaction HTML: https://canmyagentuse.com/features/automatic-context-compaction Markdown: https://canmyagentuse.com/features/automatic-context-compaction.md This row requires a documented mechanism that summarizes, prunes, or otherwise reduces earlier context so work can continue. Silent truncation without a continuity mechanism is not compaction. Evidence should record the trigger, whether the operator is warned, what state is preserved, whether instructions and tool results are treated specially, and whether the compacted summary is inspectable. Reliability tests should probe commitments, file paths, pending tasks, citations, approvals, and negative constraints across the boundary. --- ## Background agents HTML: https://canmyagentuse.com/features/background-agents Markdown: https://canmyagentuse.com/features/background-agents.md Keep a run going after the operator leaves the session. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Browser automation HTML: https://canmyagentuse.com/features/browser-automation Markdown: https://canmyagentuse.com/features/browser-automation.md Drive a browser for clicks, forms, and navigation. The sourced OpenWork cell is limited to its first-party browser. Grok Bot instead drives a browser on its hosted computer and uses an explicit takeover path for sensitive steps. Neither cell silently generalizes to the operator's local desktop or to unrestricted access across every site. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Code execution HTML: https://canmyagentuse.com/features/code-execution Markdown: https://canmyagentuse.com/features/code-execution.md Run generated code in a sandbox or constrained project environment. OpenWork's cell stays partial because the reviewed source establishes local execution and approval boundaries, not a default isolation guarantee. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Collaboration and portability HTML: https://canmyagentuse.com/features/collaboration-and-portability Markdown: https://canmyagentuse.com/features/collaboration-and-portability.md A public read-only link, a shared editable project, a queue that transfers responsibility, and a complete machine-readable export solve different problems. This family keeps those capabilities separate. Useful evidence records identity, view and edit rights, ownership, revocation, external access, inherited data permissions, comments and approvals, provenance, export fidelity, deletion behavior, and whether recipients need a paid account or matching organization. --- ## Computer use HTML: https://canmyagentuse.com/features/computer-use Markdown: https://canmyagentuse.com/features/computer-use.md This row asks whether the exact harness can inspect a rendered visual state and perform direct input actions. It distinguishes broad computer or GUI control from a browser-only feedback loop. That distinction matters: a coding harness can fully automate its integrated browser while leaving desktop applications and operating-system controls outside scope. Plan, region, policy, and runtime boundaries remain attached to each sourced cell. --- ## Concurrent subagents HTML: https://canmyagentuse.com/features/subagent-concurrency Markdown: https://canmyagentuse.com/features/subagent-concurrency.md This row requires two or more child-agent executions whose active intervals overlap. Sequential delegation, concurrent ordinary tool calls, and detached shell commands do not establish concurrent subagents. Evidence should record maximum active and queued children, nesting effects, per-plan or per-organization quotas, scheduling order, cancellation behavior, rate-limit sharing, and whether concurrent children can safely write to the same workspace. Tests should detect accidental serialization as well as unsafe shared-state races. --- ## Connectors HTML: https://canmyagentuse.com/features/connectors Markdown: https://canmyagentuse.com/features/connectors.md Reach mail, chat, files, calendars, issue trackers, and other services the user authorizes. The Grok Web and Grok Bot desktop paths are separate. A connector's appearance in a catalog does not establish that every account can authenticate it, that every exposed action is writable, or that organization policy permits it. All other harness cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Context compaction controls HTML: https://canmyagentuse.com/features/context-compaction-controls Markdown: https://canmyagentuse.com/features/context-compaction-controls.md This row separates controllability from the existence of automatic compaction. Useful controls include a manual compact action, an editable instruction for the summary, threshold configuration, inspection of the compacted state, pinning of durable facts, and an opt-out. A single manual “summarize this chat” prompt is not necessarily a harness-level compaction control. Evidence should show that the resulting state is used for subsequent requests and identify the scope of each setting. --- ## Context usage visibility HTML: https://canmyagentuse.com/features/context-usage-visibility Markdown: https://canmyagentuse.com/features/context-usage-visibility.md This row asks whether the exact harness shows the operator how much context is in use or remains. An exact token count, a percentage estimate, and a warning shown only at exhaustion are materially different and should be qualified. The most useful implementations identify contributions from conversation turns, project instructions, tool schemas, retrieved files, images or video, long-term memory, and reserved output. Evidence should also say whether the meter updates before a request, after server-side processing, or only in logs. --- ## Context window and effective input limit HTML: https://canmyagentuse.com/features/context-window Markdown: https://canmyagentuse.com/features/context-window.md “Context length” is not one provider-wide number. The catalog records the exact harness, selected model or mode, advertised total context window, maximum accepted input when separately documented, reserved output budget, and observation date. Tool definitions, hidden instructions, memory, file extraction, images, video frames, and previous turns may all consume the same budget. A sourced cell should carry the numeric limit and unit in its evidence note or structured qualifier. Mark it partial when the harness supports the model but enforces a lower cap, exposes only a qualitative label, or dynamically routes among models with different limits. Do not infer a web, desktop, editor, or CLI limit from an API model card unless the harness documentation explicitly adopts it. Useful conformance work tests boundary behavior: rejection, truncation, silent omission, compaction, retrieval, and “lost in the middle” quality are different outcomes. A large accepted input is not proof that every supplied token remains equally influential. --- ## Controlled conversation sharing HTML: https://canmyagentuse.com/features/conversation-sharing Markdown: https://canmyagentuse.com/features/conversation-sharing.md This row asks whether a conversation or run can be deliberately shared and what the recipient receives. Public anonymous links, organization-only access, named invitations, snapshots, and live collaborative views have different risk and utility. Evidence should record view, comment, edit, continue, copy, and reshare rights; revocation; expiration; search indexing; recipient account requirements; sensitive-content warnings; redaction; files and tool output; later message updates; deletion; and organization policy controls. --- ## Conversation and run export HTML: https://canmyagentuse.com/features/conversation-export Markdown: https://canmyagentuse.com/features/conversation-export.md This row asks whether an operator can take a useful run record out of the harness. Copying rendered chat text is narrower than an export containing roles, timestamps, selected and routed models, instructions, tool calls and results, approvals, citations, attachments, sub-agent relationships, errors, usage, and completion status. Evidence should record formats, schema and versioning, single and bulk export, administrator APIs, asynchronous availability, file and artifact references, redaction, signatures or checksums, retention window, and whether the data can be re-imported or replayed elsewhere. --- ## Custom instructions HTML: https://canmyagentuse.com/features/custom-instructions Markdown: https://canmyagentuse.com/features/custom-instructions.md Persist project or user instruction files. OpenWork's sourced cell covers its server-owned OpenCode workspace configuration; it does not claim that every instruction format or precedence rule is unique to OpenWork. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Data residency and processing location controls HTML: https://canmyagentuse.com/features/data-residency Markdown: https://canmyagentuse.com/features/data-residency.md This row asks where each data class is stored and processed. Storage residency does not necessarily constrain model inference, transient processing, support access, telemetry, safety systems, caches, backups, connectors, or third-party tools. Evidence should name selectable regions, default placement, migration behavior, plan and contract requirements, excluded features, disaster-recovery locations, subprocessors, and whether local clients send content outside the selected boundary. --- ## Data retention and deletion controls HTML: https://canmyagentuse.com/features/data-retention-controls Markdown: https://canmyagentuse.com/features/data-retention-controls.md This row asks what is retained, for how long, by whom, and how deletion works. The visible conversation is only one object; prompts, uploads, tool results, memory, generated files, feedback, telemetry, safety logs, caches, search indexes, and backups can have different lifetimes. Evidence should record default and configurable periods, plan and workspace scope, deletion initiation and completion timelines, administrator versus user controls, legal or abuse-monitoring exceptions, backup expiry, account-deletion behavior, and APIs or exports that let an organization verify completion. --- ## Data security and administrative controls HTML: https://canmyagentuse.com/features/data-security-controls Markdown: https://canmyagentuse.com/features/data-security-controls.md Provider-level security pages are not automatically harness-level evidence. A useful record follows prompts, uploaded files, repository content, retrieved data, tool results, generated artifacts, telemetry, feedback, abuse-monitoring copies, and backups through the exact web, desktop, editor, or CLI execution path. Each child row keeps plan, organization setting, region, account type, connector, model endpoint, local or hosted execution, and review date attached. “Enterprise ready” and a list of compliance certifications are not substitutes for operational controls an evaluator can verify. --- ## Diffs and patches HTML: https://canmyagentuse.com/features/diffs-and-patches Markdown: https://canmyagentuse.com/features/diffs-and-patches.md Propose and apply file diffs without a full git flow. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Encryption and key controls HTML: https://canmyagentuse.com/features/encryption-key-controls Markdown: https://canmyagentuse.com/features/encryption-key-controls.md This row records the cryptographic controls that apply to the exact harness path. Baseline TLS and provider-managed storage encryption are narrower than customer-managed keys, but even customer keys may exclude logs, indexes, memory, backups, connector data, or third-party tools. Evidence should identify algorithms or platform guarantees when documented, key ownership and hierarchy, tenant scope, rotation and revocation, service behavior after revocation, recovery, regional restrictions, and plan requirements. Marketing shorthand such as “end-to-end encrypted” needs a precise endpoint and key-holder definition. --- ## Failure recovery and safe retry HTML: https://canmyagentuse.com/features/failure-recovery Markdown: https://canmyagentuse.com/features/failure-recovery.md This row asks how the harness handles failures after work has begun. Blindly retrying a read is different from repeating an email send, database write, deployment, purchase, or destructive command. Safe recovery needs typed errors, bounded retry policy, idempotency or reconciliation, and operator-visible state. Evidence should record automatic and manual retries, backoff, model fallback, checkpoints, filesystem or transaction rollback, partial artifacts, tool-specific policies, duplicate prevention, child-agent failures, and whether the operator can continue from the failed step instead of restarting everything. --- ## File and media inputs HTML: https://canmyagentuse.com/features/file-inputs Markdown: https://canmyagentuse.com/features/file-inputs.md A paperclip or file picker proves very little by itself. The harness may store a file without sending its content to the model, extract only text, sample a subset of pages or frames, route some formats through a tool, or impose limits that change by plan and model. The child rows distinguish formats and record the useful operational detail: extensions and MIME types, per-file and aggregate size, file count, page or duration caps, preprocessing, resolution, transcription, archive handling, retention, and whether the original or a derivative reaches tools and sub-agents. Workspace file access is separate because a coding harness may read repository files without supporting interactive uploads, while a web chat may accept attachments without exposing a filesystem. --- ## File upload limits and quota visibility HTML: https://canmyagentuse.com/features/upload-limits Markdown: https://canmyagentuse.com/features/upload-limits.md File support is not operationally useful without its limits. This row records per-file bytes, files per message or run, pages per document, audio or video duration, image dimensions, project storage, rolling daily or hourly quotas, and any separate organization cap. Every number needs a scope: format, plan, model, client, workspace, region, and review date. Evidence should also describe the failure mode—preflight rejection, asynchronous processing failure, truncation, sampling, or silent omission—and whether the harness shows remaining quota before an upload. --- ## Git and pull requests HTML: https://canmyagentuse.com/features/git-pull-requests Markdown: https://canmyagentuse.com/features/git-pull-requests.md Inspect diffs and open or review pull requests. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Human approval HTML: https://canmyagentuse.com/features/human-approval Markdown: https://canmyagentuse.com/features/human-approval.md Pause before a tool or write until a person confirms. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Image upload and paste HTML: https://canmyagentuse.com/features/image-input Markdown: https://canmyagentuse.com/features/image-input.md This row asks whether an image selected, dragged, or pasted into the exact harness is interpreted as visual model context. Merely uploading the file to storage, extracting only its filename, or making it available to an unrelated tool does not establish image input. Evidence should record accepted formats, per-file and per-message limits, animation handling, resolution or detail controls, whether metadata is stripped, and whether availability changes by model, plan, or client. Screenshot capture is a separate capability because it concerns acquiring the current interface rather than supplying an existing image. The Grok Web and Grok Bot desktop claims are scoped independently because their attachment limits and runtime behavior differ. Image input does not by itself establish screenshot capture, computer use, image generation, or reliable interpretation of every image format. All other harness cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Local models HTML: https://canmyagentuse.com/features/local-models Markdown: https://canmyagentuse.com/features/local-models.md This row asks whether the exact harness can send model requests to an operator-controlled local inference runtime. A cloud model with a desktop client is not a local model, and a configurable remote API base URL does not prove on-device execution. Evidence should identify the supported runtimes or protocols, model-discovery behavior, tool-calling and multimodal constraints, authentication, and whether any prompts, telemetry, or embeddings still leave the device. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Local-only and offline operation HTML: https://canmyagentuse.com/features/local-only-mode Markdown: https://canmyagentuse.com/features/local-only-mode.md A native desktop or CLI process is not automatically local-only. This row requires a documented workflow in which model inference, required tools, storage, authentication, policy evaluation, and necessary control-plane functions remain available without sending task content to a hosted service. Evidence should record initial and recurring sign-in, license checks, model downloads, updates, telemetry, crash reports, remote connectors, disabled features, offline duration, DNS and network behavior, and the steps required for an air-gapped installation. Partial support is appropriate when the core run is local but setup or optional features require the cloud. --- ## Long-term memory HTML: https://canmyagentuse.com/features/long-term-memory Markdown: https://canmyagentuse.com/features/long-term-memory.md Persist preferences and working context across sessions. The sourced Grok Bot cell covers the documented memory of a named Bot. It does not imply perfect recall, a user-visible memory editor, isolation between Bots, or that remembered facts are current. All other harness cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## MCP Apps HTML: https://canmyagentuse.com/features/mcp-apps Markdown: https://canmyagentuse.com/features/mcp-apps.md MCP Apps is the official Model Context Protocol extension for interactive interfaces returned by tools. A compatible host fetches a declared UI resource, renders it in a sandboxed frame, and mediates communication between the interface and MCP tools. This row is narrower than ordinary MCP tool calling. A harness can support MCP tools while leaving this UI extension unknown. The first wave sources only hosts named by current product or protocol documentation. --- ## MCP cancellation HTML: https://canmyagentuse.com/features/mcp-cancellation Markdown: https://canmyagentuse.com/features/mcp-cancellation.md Cancel an in-flight MCP request and handle cancellation correctly. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP client role HTML: https://canmyagentuse.com/features/mcp-client-role Markdown: https://canmyagentuse.com/features/mcp-client-role.md Consume capabilities exposed by MCP servers. The Grok Bot claim is scoped to its hosted desktop product and inherited Cursor MCP policy. It does not prove support for MCP resources, prompts, sampling, elicitation, or every authentication flow; those remain separate rows. --- ## MCP elicitation HTML: https://canmyagentuse.com/features/mcp-elicitation Markdown: https://canmyagentuse.com/features/mcp-elicitation.md Handle structured user-input requests initiated by an MCP server. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP legacy HTTP and SSE HTML: https://canmyagentuse.com/features/mcp-legacy-sse Markdown: https://canmyagentuse.com/features/mcp-legacy-sse.md Connect through the superseded HTTP plus SSE transport when migration compatibility is required. The Grok Bot cell is registry-listed for migration compatibility. New deployments should prefer the current transport where the server supports it. --- ## MCP list-changed notifications HTML: https://canmyagentuse.com/features/mcp-list-changed Markdown: https://canmyagentuse.com/features/mcp-list-changed.md Refresh capability lists when a server announces dynamic changes. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP logging HTML: https://canmyagentuse.com/features/mcp-logging Markdown: https://canmyagentuse.com/features/mcp-logging.md Receive structured log messages emitted by an MCP server. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP OAuth authorization HTML: https://canmyagentuse.com/features/mcp-oauth Markdown: https://canmyagentuse.com/features/mcp-oauth.md Complete the MCP authorization flow for protected remote servers. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP progress notifications HTML: https://canmyagentuse.com/features/mcp-progress Markdown: https://canmyagentuse.com/features/mcp-progress.md Receive progress updates for long-running requests. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP prompt-list notifications HTML: https://canmyagentuse.com/features/mcp-prompts-list-changed Markdown: https://canmyagentuse.com/features/mcp-prompts-list-changed.md Refresh prompt templates after a server announces changes. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP prompts HTML: https://canmyagentuse.com/features/mcp-prompts Markdown: https://canmyagentuse.com/features/mcp-prompts.md Invoke reusable MCP prompt templates. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## MCP Registry metadata HTML: https://canmyagentuse.com/features/mcp-registry Markdown: https://canmyagentuse.com/features/mcp-registry.md Discover MCP servers and metadata through the official registry model. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP resource subscriptions HTML: https://canmyagentuse.com/features/mcp-resource-subscriptions Markdown: https://canmyagentuse.com/features/mcp-resource-subscriptions.md Subscribe to resource changes and receive update notifications. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP resources HTML: https://canmyagentuse.com/features/mcp-resources Markdown: https://canmyagentuse.com/features/mcp-resources.md Read MCP resources as live context. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## MCP revision compatibility HTML: https://canmyagentuse.com/features/mcp-revision Markdown: https://canmyagentuse.com/features/mcp-revision.md Implement or negotiate the exact dated MCP revision. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP roots HTML: https://canmyagentuse.com/features/mcp-roots Markdown: https://canmyagentuse.com/features/mcp-roots.md Provide filesystem or workspace roots to an MCP server. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP sampling HTML: https://canmyagentuse.com/features/mcp-sampling Markdown: https://canmyagentuse.com/features/mcp-sampling.md Handle server requests for model sampling through the client. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP server instructions HTML: https://canmyagentuse.com/features/mcp-server-instructions Markdown: https://canmyagentuse.com/features/mcp-server-instructions.md Consume server-provided instructions during initialization. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP server role HTML: https://canmyagentuse.com/features/mcp-server-role Markdown: https://canmyagentuse.com/features/mcp-server-role.md Expose capabilities from the harness as an MCP server. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP stdio transport HTML: https://canmyagentuse.com/features/mcp-stdio Markdown: https://canmyagentuse.com/features/mcp-stdio.md Connect to a local MCP server over standard input and output. The Grok Bot cell is registry-listed. Server lifecycle, installation location, authorization, and security policy remain deployment-specific. --- ## MCP Streamable HTTP HTML: https://canmyagentuse.com/features/mcp-streamable-http Markdown: https://canmyagentuse.com/features/mcp-streamable-http.md Connect to an MCP server over the current Streamable HTTP transport. The Grok Bot cell is registry-listed. Authentication, server policy, and optional MCP primitives remain separate claims. --- ## MCP task listing HTML: https://canmyagentuse.com/features/mcp-task-listing Markdown: https://canmyagentuse.com/features/mcp-task-listing.md List and page through tasks exposed by the task utility. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP tasks HTML: https://canmyagentuse.com/features/mcp-tasks Markdown: https://canmyagentuse.com/features/mcp-tasks.md Create, inspect, and manage long-running MCP tasks. This row remains unknown for a product until exact, dated evidence covers this capability rather than MCP generally. --- ## MCP tools HTML: https://canmyagentuse.com/features/mcp-tools Markdown: https://canmyagentuse.com/features/mcp-tools.md MCP tools are operations advertised by a Model Context Protocol server and made available for a model or agent to invoke. This row asks whether the named harness can connect to an MCP server and call its tools. It does not claim support for every MCP transport, authentication flow, protocol primitive, or third-party server. Plan, rollout, and transport restrictions are marked partial when the public documentation states them. Only the current track is sourced in this first wave. Preview and unreviewed harness cells remain unknown. --- ## Model Context Protocol HTML: https://canmyagentuse.com/features/mcp Markdown: https://canmyagentuse.com/features/mcp.md MCP support is not one switch. A product may invoke server tools while it does not expose server prompts, read resources, or render MCP Apps. This family page derives its progress from the independently evidenced child capabilities. Open a child row for exact targets, constraints, dates, and evidence. --- ## Model fallback visibility and control HTML: https://canmyagentuse.com/features/model-fallback Markdown: https://canmyagentuse.com/features/model-fallback.md This row tracks whether a harness may retry or continue with a different model and what control the operator has over that change. Transparent, configurable fallback is distinct from silent automatic routing. Evidence should record triggers, candidate models, changes in context or tool compatibility, whether the operator is notified, how provenance appears in logs or exports, and whether regulated or reproducible workflows can disable fallback. --- ## Model selection HTML: https://canmyagentuse.com/features/model-selection Markdown: https://canmyagentuse.com/features/model-selection.md This row asks whether the operator can select a specific model or a clearly documented model class for the exact harness and task. An opaque automatic router is not explicit selection; a “fast” or “deep” mode may be partial if its underlying model can change without notice. Evidence should record eligible plans, model aliases and revisions, per-workspace restrictions, whether tools and media inputs narrow the choices, and whether the selection persists for a conversation, agent, project, or single run. --- ## Model training and improvement controls HTML: https://canmyagentuse.com/features/training-data-controls Markdown: https://canmyagentuse.com/features/training-data-controls.md This row separates several uses that are often compressed into “not used for training”: foundation-model training, service improvement, evaluation, personalization, abuse prevention, safety review, human review, and content a user intentionally submits as feedback. Evidence should identify defaults by account and plan, opt-in or opt-out mechanics, administrator enforcement, whether connectors and tool output follow the same rule, policy-change notice, and whether third-party model or tool providers receive content under different terms. --- ## Models and context HTML: https://canmyagentuse.com/features/models-and-context Markdown: https://canmyagentuse.com/features/models-and-context.md A harness is not a transparent pass-through to a model. It can select or route models, reserve tokens for hidden instructions and tools, enforce a lower input cap, compact earlier turns, retry on another model, and cache only certain prompt segments. This family therefore avoids copying a provider model-card number into every product surface. Each child row asks a narrower question and keeps model, mode, plan, workspace policy, file-processing cost, tool-schema overhead, and observation date attached to the answer. For context limits, useful records report the advertised maximum and the effective usable budget when those differ. For caching, they distinguish reuse itself from explicit controls and operator-visible telemetry. --- ## Multi-file edit HTML: https://canmyagentuse.com/features/multi-file-edit Markdown: https://canmyagentuse.com/features/multi-file-edit.md Change several files in one reviewed step. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Native plugin system HTML: https://canmyagentuse.com/features/native-plugins Markdown: https://canmyagentuse.com/features/native-plugins.md This capability covers the product's own extension mechanism. It does not establish portability to another host. --- ## Nested AGENTS.md instructions HTML: https://canmyagentuse.com/features/nested-agents-md Markdown: https://canmyagentuse.com/features/nested-agents-md.md Apply nested AGENTS.md files with directory-sensitive precedence. This row remains unknown until exact, dated evidence covers this behavior. --- ## Nested subagents HTML: https://canmyagentuse.com/features/nested-subagents Markdown: https://canmyagentuse.com/features/nested-subagents.md Allow a child agent to delegate further work to another child agent. This row remains unknown until exact, dated evidence covers this behavior. --- ## Output token limit HTML: https://canmyagentuse.com/features/output-token-limit Markdown: https://canmyagentuse.com/features/output-token-limit.md This row tracks the maximum response or artifact generation budget exposed by the harness. The number may be lower than a model API limit and may vary for chat, code edits, tool calls, structured output, or background tasks. Evidence should identify units, model and mode, defaults, user controls, stop behavior, and whether the harness continues long results across messages or files. UI clipping, transport truncation, and model generation limits should not be collapsed into one cause. --- ## PDF documents HTML: https://canmyagentuse.com/features/pdf-documents Markdown: https://canmyagentuse.com/features/pdf-documents.md This row asks whether the exact harness can accept a PDF and expose useful document content to the model. Evidence should distinguish text extraction from page-image understanding, tables and forms from plain text, scanned PDFs from born-digital files, and password-protected files from ordinary documents. Record page, file-size, file-count, plan, and model limits as qualifiers. A generic file picker or workspace file reader does not by itself prove PDF parsing. The Grok surfaces remain separate: Grok Web has a broader documented attachment count, while Grok Bot's desktop composer accepts up to six attachments and limits each document to 25 MB. These cells establish PDF input, not reliable OCR for every scan, support for encrypted files, or identical handling of embedded media. All other harness cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Per-subagent model selection HTML: https://canmyagentuse.com/features/subagent-model-selection Markdown: https://canmyagentuse.com/features/subagent-model-selection.md This row requires a per-child model selection, role default, or enforceable model constraint. A model chosen once for the whole conversation does not establish independent sub-agent choice. Evidence should record which actors can select the model, allowed models and reasoning modes, fallbacks, plan restrictions, tool and media compatibility, cost attribution, and whether nested agents inherit or override the selection. --- ## Persistent instructions HTML: https://canmyagentuse.com/features/instructions Markdown: https://canmyagentuse.com/features/instructions.md Reading one root instruction file does not establish nested precedence, path-specific rules, or compatibility with a product-native instruction format. --- ## Plan and pricing disclosure HTML: https://canmyagentuse.com/features/plan-pricing-disclosure Markdown: https://canmyagentuse.com/features/plan-pricing-disclosure.md This row asks whether a prospective operator can determine what the exact harness costs and which plan unlocks the evaluated behavior. A headline monthly price is incomplete when model premiums, tool use, hosted compute, storage, connectors, sub-agents, or overages are billed separately. Evidence should record billing unit and period, included quota, variable rates, hard and soft caps, seat minimums, annual commitments, currency and tax treatment, trial behavior, enterprise-only terms, and the effective date. The catalog does not convert those facts into a synthetic “value” score. --- ## Plugin marketplace and installation HTML: https://canmyagentuse.com/features/plugin-marketplace Markdown: https://canmyagentuse.com/features/plugin-marketplace.md Discover, install, update, and remove extensions through a registry or marketplace. The Grok Bot cell establishes its in-product discovery and installation flow. It does not imply that every listed plugin is available to every plan or organization, that all plugins use one portable format, or that Can My Agent Use reviewed third-party packages. --- ## Plugin systems HTML: https://canmyagentuse.com/features/plugins Markdown: https://canmyagentuse.com/features/plugins.md “Supports plugins” is not a sufficient compatibility claim. A product-native extension, a portable Agent Plugins package, a packaged skill, and a marketplace install can have different formats, trust boundaries, and lifecycle behavior. This family derives progress from independently evidenced atomic capabilities. --- ## Plugin vendor extensions HTML: https://canmyagentuse.com/features/plugin-vendor-extensions Markdown: https://canmyagentuse.com/features/plugin-vendor-extensions.md Preserve namespaced host-specific behavior without representing it as portable core support. This row remains unknown until exact, dated evidence covers this behavior. --- ## Product-native rules HTML: https://canmyagentuse.com/features/native-rules Markdown: https://canmyagentuse.com/features/native-rules.md Discover persistent or path-scoped instructions in the product's own rule format. This row remains unknown until exact, dated evidence covers this behavior. --- ## Project-scoped Agent Skills HTML: https://canmyagentuse.com/features/skill-project-scope Markdown: https://canmyagentuse.com/features/skill-project-scope.md Discover skills installed for a repository or workspace. This row remains unknown until exact, dated evidence covers this behavior. --- ## Prompt and context cache reuse HTML: https://canmyagentuse.com/features/prompt-cache-reuse Markdown: https://canmyagentuse.com/features/prompt-cache-reuse.md This row concerns model-request prompt or context caching: eligible repeated prefixes receive documented processing, latency, or billing reuse. It does not include browser caches, downloaded-file caches, embedding indexes, retrieval caches, build caches, or a conversation merely retaining its history. Evidence should identify whether reuse is automatic or explicit, minimum eligible prefix size, exact-prefix requirements, supported models and regions, cache lifetime, isolation boundary, invalidation behavior, and read-versus-write pricing. A model API feature does not prove that a hosted chat or coding harness preserves stable prefixes or passes cache controls through. --- ## Prompt cache controls HTML: https://canmyagentuse.com/features/prompt-cache-controls Markdown: https://canmyagentuse.com/features/prompt-cache-controls.md This row asks whether the harness exposes a supported way to influence prompt caching rather than relying entirely on automatic provider behavior. Controls can include cache breakpoints, explicit cache objects, retention choices, refresh, or invalidation. Evidence should record where the control lives—UI, project configuration, command flag, plugin, or provider-specific request setting—and whether the exact harness preserves it. Security review should also identify the tenant, workspace, user, project, or organization boundary around reuse. --- ## Prompt cache telemetry HTML: https://canmyagentuse.com/features/prompt-cache-telemetry Markdown: https://canmyagentuse.com/features/prompt-cache-telemetry.md This row asks whether an operator can verify prompt-cache behavior for the exact harness. A billing dashboard total, per-request cached-token count, cache-hit event, and detailed read/write breakdown provide different levels of evidence and should be qualified. Useful records identify where telemetry appears, how long it is retained, whether sub-agent and tool-generated requests are included, and how cached reads and writes map to price. A generic statement that caching is automatic does not prove that the harness exposes telemetry. --- ## Rate limit and quota disclosure HTML: https://canmyagentuse.com/features/rate-limit-disclosure Markdown: https://canmyagentuse.com/features/rate-limit-disclosure.md This row records rate and quota boundaries around the harness: requests, tokens, agent runs, active sandboxes, sub-agent concurrency, tool calls, connectors, uploads, media minutes, storage, and background compute. Each number needs a plan, model, user or organization scope, and reset window. Evidence should state whether pools are shared, whether limits are fixed or dynamic, what meters or headers reveal, retry and backoff guidance, priority tiers, overage behavior, and whether work queues, degrades, or fails at the boundary. “Reasonable use” without an inspectable limit is partial disclosure. --- ## Realtime voice HTML: https://canmyagentuse.com/features/realtime-voice Markdown: https://canmyagentuse.com/features/realtime-voice.md Speak and listen over a live audio session. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Reasoning effort controls HTML: https://canmyagentuse.com/features/reasoning-effort-controls Markdown: https://canmyagentuse.com/features/reasoning-effort-controls.md This row covers an operator-visible control over reasoning effort, depth, or budget. It does not assert that hidden reasoning is exposed, and it does not treat ordinary model selection as an equivalent control. Evidence should name the available levels, defaults, model and plan restrictions, whether the setting changes output-token or tool-use budgets, and the scope at which it persists. Qualify controls that are only advisory or available through configuration rather than the primary interface. --- ## Resumable runs HTML: https://canmyagentuse.com/features/resumable-runs Markdown: https://canmyagentuse.com/features/resumable-runs.md This row requires more than viewing an old transcript. The harness must preserve enough execution state to reconnect to an active run or continue from a checkpoint after closure, network loss, process restart, pause, or worker migration. Evidence should identify which state survives: conversation and instructions, tool results, filesystem and git state, browser sessions, pending approvals, sub-agents, environment variables, secrets, timers, and streamed output. Record resume window, idempotency, billing, cross-device behavior, and cases that force a fresh run. --- ## Role-based collaboration access HTML: https://canmyagentuse.com/features/role-based-access Markdown: https://canmyagentuse.com/features/role-based-access.md This row asks whether collaboration rights are enforced by role rather than inferred from possession of a link. Useful permission boundaries separate viewing prompts and files, editing context, starting runs, using billable models, invoking tools, approving actions, managing connectors or secrets, sharing externally, exporting, deleting, and administering policy. Evidence should record built-in and custom roles, project and organization scope, group mapping, temporary access, inheritance and exceptions, removal behavior for active runs, service identities, and audit attribution. --- ## Run timeout and duration limits HTML: https://canmyagentuse.com/features/run-timeout-disclosure Markdown: https://canmyagentuse.com/features/run-timeout-disclosure.md One “timeout” can conceal several clocks: total wall time, active compute, inactivity, tool-call duration, sandbox lifetime, browser session, approval wait, connector deadline, and child-agent or nested-agent limits. Evidence should record each limit, plan and environment, whether activity resets the clock, warning behavior, checkpointing, partial artifacts, cancellation propagation, billing after the client disconnects, and cleanup. Observed timeouts are useful test evidence but should not be mislabeled as documented guarantees. --- ## Run, approval, and failure notifications HTML: https://canmyagentuse.com/features/run-notifications Markdown: https://canmyagentuse.com/features/run-notifications.md This row asks whether the harness can tell the responsible operator that unattended work completed, failed, stalled, hit a budget or quota, or needs approval. An in-app badge visible only after reopening the product is narrower than an external actionable notification. Evidence should record supported channels, event types, user and project routing, quiet hours, escalation, deduplication, links back to the exact run, sensitive-content redaction, access checks, webhook signing, retry behavior, and notifications generated by sub-agents. --- ## Sandbox network HTML: https://canmyagentuse.com/features/sandbox-network Markdown: https://canmyagentuse.com/features/sandbox-network.md Allow or deny outbound network from a tool sandbox. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Scheduled runs HTML: https://canmyagentuse.com/features/scheduled-runs Markdown: https://canmyagentuse.com/features/scheduled-runs.md Start a task on a timer without a human prompt. Grok Bot's sourced cell covers product-native routines owned by one Bot. It does not establish unlimited unattended operation, guaranteed event delivery, or identical retention for the conversation and run history. All other harness cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Screenshots HTML: https://canmyagentuse.com/features/screenshots Markdown: https://canmyagentuse.com/features/screenshots.md Capture the current screen or window as context. OpenWork's sourced cell is deliberately partial because its documented screenshot path is limited to the first-party browser. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Secrets management and redaction HTML: https://canmyagentuse.com/features/secrets-management Markdown: https://canmyagentuse.com/features/secrets-management.md This row asks how a harness handles credentials used by tools, connectors, repositories, terminals, MCP servers, and sub-agents. Plain environment variables or pasted API keys are not equivalent to a managed secret that can be injected without revealing its value to model context or logs. Evidence should record storage boundary, encryption, user and workspace scope, tool allowlists, just-in-time grants, masking, rotation, revocation, export prevention, prompt-injection defenses, child-agent inheritance, and audit events. Redaction is best-effort unless the documented mechanism prevents the model from receiving the value in the first place. --- ## Shared projects and team context HTML: https://canmyagentuse.com/features/shared-projects Markdown: https://canmyagentuse.com/features/shared-projects.md This row requires a durable collaboration boundary that can hold more than an isolated shared chat. A project may include members and roles, instructions, knowledge files, memory, connectors, saved agents, conversations, artifacts, budgets, and policy. Evidence should record which objects are shared, ownership and transfer, invitation and removal, inherited connector permissions, model and tool policy, version history, data isolation between projects, export and deletion, and how personal context is prevented from leaking into team work. --- ## Streaming output HTML: https://canmyagentuse.com/features/streaming-output Markdown: https://canmyagentuse.com/features/streaming-output.md Show tokens and tool events as they arrive. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Structured output HTML: https://canmyagentuse.com/features/structured-output Markdown: https://canmyagentuse.com/features/structured-output.md Return JSON or schema-validated objects. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## Structured subagent result handoff HTML: https://canmyagentuse.com/features/subagent-result-handoff Markdown: https://canmyagentuse.com/features/subagent-result-handoff.md This row asks what crosses the child-to-parent boundary. A plain prose summary may be useful, but it can lose status, sources, files, patches, tool output, uncertainty, and failure details. Structured handoff preserves enough provenance for the parent and operator to verify and act on the result. Evidence should identify supported result types, size limits, streaming versus final delivery, partial results after cancellation, error representation, duplicate or late completion handling, and whether created artifacts remain addressable outside the child's private workspace. --- ## Subagent approval and permission boundaries HTML: https://canmyagentuse.com/features/subagent-approval-boundaries Markdown: https://canmyagentuse.com/features/subagent-approval-boundaries.md This row asks how authority crosses a delegation boundary. A child may inherit every parent grant, receive a reduced tool set, prompt the operator directly, ask the parent to broker an action, or be unable to escalate. Each design changes accountability and prompt-injection risk. Evidence should cover one-time and persistent approvals, resource and command scope, MCP and connector credentials, filesystem and network access, nested descendants, denial propagation, audit identity, and whether the operator can review a combined action plan before fan-out. --- ## Subagent context isolation and inheritance HTML: https://canmyagentuse.com/features/subagent-context-isolation Markdown: https://canmyagentuse.com/features/subagent-context-isolation.md This row asks whether the child-agent boundary has documented context semantics. A child may start fresh, receive the full parent transcript, inherit only a task prompt and selected files, share project memory, or receive a compacted summary. Those modes have different correctness, privacy, and token-cost consequences. Evidence should record inherited system and project instructions, conversation turns, tool schemas, environment variables, files, long-term memory, credentials, approval state, and context budget. It should also say whether later parent changes propagate and whether the operator can inspect the exact handoff. --- ## Subagent delegation HTML: https://canmyagentuse.com/features/subagent-delegation Markdown: https://canmyagentuse.com/features/subagent-delegation.md This row establishes that one agent can create a distinct child execution boundary for a bounded task. A renamed tool call, background shell command, or single-model chain-of-thought step is not enough unless the product exposes a child-agent execution with its own task state. Deeper recursion, simultaneous children, inherited context, model choice, permissions, supervision, and result transfer are separate sibling capabilities. --- ## Subagent lifecycle controls HTML: https://canmyagentuse.com/features/subagent-lifecycle-controls Markdown: https://canmyagentuse.com/features/subagent-lifecycle-controls.md This row tracks supervision of child-agent executions after delegation. At minimum, evidence should show how active and queued children are identified and whether cancellation propagates to their tools and descendants. Record pause, resume, retry, deadline, timeout, priority, log inspection, partial-result recovery, and orphan cleanup separately. A UI that hides a child after cancel while its tools continue running is only partial support. --- ## Subagent MCP access HTML: https://canmyagentuse.com/features/subagent-mcp-access Markdown: https://canmyagentuse.com/features/subagent-mcp-access.md Allow a child agent to use MCP tools or declare its own MCP servers. This row remains unknown until exact, dated evidence covers this behavior. --- ## Subagent write access HTML: https://canmyagentuse.com/features/subagent-write-access Markdown: https://canmyagentuse.com/features/subagent-write-access.md Allow a child agent to create or modify project files. This row remains unknown until exact, dated evidence covers this behavior. --- ## Subagents HTML: https://canmyagentuse.com/features/subagents Markdown: https://canmyagentuse.com/features/subagents.md A harness may delegate one bounded task while preventing parallel work, nested delegation, model selection, MCP access, project writes, or independent approvals. It may return only a text summary even when the child produced files or structured artifacts. The family rollup keeps those boundaries visible instead of promoting one successful child task into a blanket “multi-agent” claim. Useful evidence records who can create a child, how much context it inherits, whether children share mutable state, concurrency and nesting limits, selectable models, available tools, approval ownership, cancellation and inspection controls, and how results return to the parent. --- ## Terminal commands HTML: https://canmyagentuse.com/features/terminal Markdown: https://canmyagentuse.com/features/terminal.md This capability covers a built-in, permissioned route for an agent to propose or execute shell commands in its working environment. It does not claim unattended execution, access beyond the selected workspace, a particular shell, or identical approval behavior. Only current product documentation is used here. Preview tracks and unreviewed surfaces remain unknown. --- ## Text and office document input HTML: https://canmyagentuse.com/features/office-document-input Markdown: https://canmyagentuse.com/features/office-document-input.md This row covers common document attachments such as plain text, rich text, word-processing files, presentations, and spreadsheets. Evidence should list exact formats and distinguish plain-text extraction from retained tables, formulas, notes, comments, charts, images, layout, and workbook or slide structure. Macros, external links, encrypted files, tracked changes, hidden sheets, and embedded objects are separate risk and fidelity boundaries. When a connector opens a cloud document by reference, record that path separately from a direct upload. --- ## Usage and cost metering HTML: https://canmyagentuse.com/features/usage-metering Markdown: https://canmyagentuse.com/features/usage-metering.md This row asks whether operators can see consumption close enough to a run to diagnose and control cost. Useful meters separate input, output, reasoning, and cached tokens; tool and connector calls; hosted sandbox or GPU time; storage; media processing; and parent versus sub-agent usage. Evidence should record update delay, estimated versus final values, price mapping, currencies, user/project/organization attribution, export API, budgets and alerts, model routing, and treatment of retries and failed runs. A monthly invoice total is only partial operational visibility. --- ## Usage, cost, and run reliability HTML: https://canmyagentuse.com/features/usage-and-reliability Markdown: https://canmyagentuse.com/features/usage-and-reliability.md A feature can be technically supported but operationally unusable for a workload because its cost is opaque, its quota is shared, its run times out, or a disconnected client loses all progress. This family makes those day-two questions visible. Useful evidence attaches every amount and limit to a plan, billing region and currency, model or mode, organization scope, reset period, client, execution environment, and review date. It distinguishes a documented maximum from an observed one and records what happens at the boundary. --- ## User-scoped Agent Skills HTML: https://canmyagentuse.com/features/skill-user-scope Markdown: https://canmyagentuse.com/features/skill-user-scope.md Discover personal or global skills outside the current project. This row remains unknown until exact, dated evidence covers this behavior. --- ## Video preprocessing disclosure HTML: https://canmyagentuse.com/features/video-processing-disclosure Markdown: https://canmyagentuse.com/features/video-processing-disclosure.md Two harnesses may both accept the same video while presenting very different information to the model. This row asks whether the product explains frame sampling, keyframe selection, scene detection, audio extraction, transcription, resizing, duration truncation, and context or quota accounting. The claim is about disclosure, not visual quality. A partial status is appropriate when only some transformations or limits are documented. Runtime testing should compare the documented pipeline with time-specific and audio-visual questions. --- ## Video upload and understanding HTML: https://canmyagentuse.com/features/video-input Markdown: https://canmyagentuse.com/features/video-input.md This row asks whether an uploaded or selected video contributes content to the model. Evidence must say what “video understanding” means for that harness: sampled frames, native temporal input, extracted audio, generated transcript, metadata, or some combination. Record accepted containers and codecs, maximum bytes and duration, frame or sampling policy, audio handling, timecode awareness, resolution changes, model and plan restrictions, processing latency, and whether links are supported in addition to local uploads. A harness that accepts a video only for storage, sharing, or an unrelated editing tool remains unsupported for this row. Testing should include questions that require temporal order, scene changes, audio-visual alignment, small on-screen text, and time-specific citations; a good transcript answer alone does not prove visual or temporal understanding. --- ## Web fetch HTML: https://canmyagentuse.com/features/web-fetch Markdown: https://canmyagentuse.com/features/web-fetch.md Retrieve live web pages or APIs during a run. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## WebMCP HTML: https://canmyagentuse.com/features/webmcp Markdown: https://canmyagentuse.com/features/webmcp.md WebMCP is not remote MCP server configuration. It is a draft browser API through which a page exposes structured tools to browser-provided or author-provided agents. The family keeps page authoring APIs, agent consumption APIs, origin controls, and security signals distinct. --- ## WebMCP declarative form tools HTML: https://canmyagentuse.com/features/webmcp-declarative-tools Markdown: https://canmyagentuse.com/features/webmcp-declarative-tools.md Expose semantic HTML forms as structured tools through WebMCP annotations. WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## WebMCP execution cancellation HTML: https://canmyagentuse.com/features/webmcp-cancellation Markdown: https://canmyagentuse.com/features/webmcp-cancellation.md Cancel pending tool registration or execution through abort signals and lifecycle rules. WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## WebMCP imperative tools HTML: https://canmyagentuse.com/features/webmcp-imperative-tools Markdown: https://canmyagentuse.com/features/webmcp-imperative-tools.md Register JavaScript-backed tools through document.modelContext.registerTool(). WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## WebMCP permissions policy HTML: https://canmyagentuse.com/features/webmcp-permissions-policy Markdown: https://canmyagentuse.com/features/webmcp-permissions-policy.md Gate tool exposure across documents and origins through the tools permissions policy. WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## WebMCP security annotations HTML: https://canmyagentuse.com/features/webmcp-security-annotations Markdown: https://canmyagentuse.com/features/webmcp-security-annotations.md Communicate read-only and untrusted-content hints across the site-to-agent boundary. WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## WebMCP tool discovery HTML: https://canmyagentuse.com/features/webmcp-tool-discovery Markdown: https://canmyagentuse.com/features/webmcp-tool-discovery.md Expose the active page tool list and schemas to a WebMCP-aware agent. WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## WebMCP tool execution HTML: https://canmyagentuse.com/features/webmcp-tool-execution Markdown: https://canmyagentuse.com/features/webmcp-tool-execution.md Invoke a page tool with structured arguments and return its result to the agent. WebMCP is a draft Community Group Report, not a W3C Standard. Positive implementation cells must retain origin-trial, version, permissions, and agent-integration qualifiers. --- ## Workspace files HTML: https://canmyagentuse.com/features/workspace-files Markdown: https://canmyagentuse.com/features/workspace-files.md This row asks whether a harness can directly read or change files inside a local project that the user has opened or selected. It does not imply unrestricted access to the whole computer, cloud drives, or files outside the tool's permission boundary. Desktop support is marked partial when file access is supplied through a separately installed extension or MCP server. The current track is the only reviewed track in this first evidence wave. --- ## Workspace search HTML: https://canmyagentuse.com/features/workspace-search Markdown: https://canmyagentuse.com/features/workspace-search.md Search a codebase or knowledge tree before editing. The matrix below lists published **web**, **desktop**, and **CLI** surfaces. Unreviewed cells remain **unknown** until a dated note and public source support a more specific status. Use the Markdown and JSON twins if you are an agent reading this site. --- ## ChatGPT (web) HTML: https://canmyagentuse.com/harnesses/chatgpt-web Markdown: https://canmyagentuse.com/harnesses/chatgpt-web.md ChatGPT appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of OpenAI and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Claude (web) HTML: https://canmyagentuse.com/harnesses/claude-web Markdown: https://canmyagentuse.com/harnesses/claude-web.md Claude appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Anthropic and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Gemini (web) HTML: https://canmyagentuse.com/harnesses/gemini-web Markdown: https://canmyagentuse.com/harnesses/gemini-web.md Gemini appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Google and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Copilot (web) HTML: https://canmyagentuse.com/harnesses/copilot-web Markdown: https://canmyagentuse.com/harnesses/copilot-web.md Copilot appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Microsoft and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Grok (web) HTML: https://canmyagentuse.com/harnesses/grok-web Markdown: https://canmyagentuse.com/harnesses/grok-web.md Grok appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of xAI and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Grok Bot (desktop) HTML: https://canmyagentuse.com/harnesses/grok-bot-desktop Markdown: https://canmyagentuse.com/harnesses/grok-bot-desktop.md Grok Bot appears here as a distinct **Desktop** harness because it is not the same surface as the consumer Grok web chat. The reviewed documentation requires the macOS or Windows app, while each Bot's browser, filesystem, terminal, tools, and durable state live on a hosted computer. Current documentation limits access to eligible SuperGrok or Cursor plans and says Linux has no desktop app. All Bots for one account share that account's cloud computer, including its files, browser sessions, and command-line credentials; separate Bots are therefore not separate security boundaries. Access to the operator's local computer is a separate, approval-controlled capability. The product names are plain-text nominative references. Can My Agent Use is independent of xAI and Cursor and does not claim certification, partnership, or official status. Capability cells cite dated public sources and retain plan, platform, runtime, and policy qualifications; unreviewed cells remain **unknown**. --- ## Perplexity (web) HTML: https://canmyagentuse.com/harnesses/perplexity-web Markdown: https://canmyagentuse.com/harnesses/perplexity-web.md Perplexity appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Perplexity and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Le Chat (web) HTML: https://canmyagentuse.com/harnesses/le-chat Markdown: https://canmyagentuse.com/harnesses/le-chat.md Le Chat appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Mistral and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Devin (web) HTML: https://canmyagentuse.com/harnesses/devin-web Markdown: https://canmyagentuse.com/harnesses/devin-web.md Devin appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Cognition and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Replit Agent (web) HTML: https://canmyagentuse.com/harnesses/replit-agent Markdown: https://canmyagentuse.com/harnesses/replit-agent.md Replit Agent appears here as a **Web** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Replit and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## ChatGPT (desktop) HTML: https://canmyagentuse.com/harnesses/chatgpt-desktop Markdown: https://canmyagentuse.com/harnesses/chatgpt-desktop.md ChatGPT appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of OpenAI and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Claude (desktop) HTML: https://canmyagentuse.com/harnesses/claude-desktop Markdown: https://canmyagentuse.com/harnesses/claude-desktop.md Claude appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Anthropic and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Cursor (desktop) HTML: https://canmyagentuse.com/harnesses/cursor Markdown: https://canmyagentuse.com/harnesses/cursor.md Cursor appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Anysphere and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## OpenWork Desktop (desktop) HTML: https://canmyagentuse.com/harnesses/openwork-desktop Markdown: https://canmyagentuse.com/harnesses/openwork-desktop.md OpenWork Desktop is the publisher's first-party **Desktop** agent harness. It runs an OpenCode-backed host locally, can connect to remote OpenWork workers, and exposes documented flows for skills, MCP servers, native OpenCode plugins, permission decisions, and a built-in browser. The catalog treats it as a full comparison surface rather than a name-only placeholder. The product name is a plain-text nominative reference; Can My Agent Use is independent of OpenWork and does not claim certification, partnership, or endorsement. Support cells for this surface are published only when a dated note cites reviewable public evidence. Platform, subscription, remote-worker, and companion-extension conditions stay attached to the relevant cells; every other cell remains **unknown**. --- ## Copilot Chat (desktop) HTML: https://canmyagentuse.com/harnesses/vscode-copilot Markdown: https://canmyagentuse.com/harnesses/vscode-copilot.md Copilot Chat appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Microsoft and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Chrome WebMCP origin trial (desktop) HTML: https://canmyagentuse.com/harnesses/chrome-webmcp-preview Markdown: https://canmyagentuse.com/harnesses/chrome-webmcp-preview.md This exact column represents the Chrome WebMCP origin trial documented by the Chrome team. It does not claim that WebMCP is enabled by default in stable Chrome or available to every Gemini in Chrome user. --- ## Windsurf (desktop) HTML: https://canmyagentuse.com/harnesses/windsurf Markdown: https://canmyagentuse.com/harnesses/windsurf.md Windsurf appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Codeium and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Zed Agent (desktop) HTML: https://canmyagentuse.com/harnesses/zed-agent Markdown: https://canmyagentuse.com/harnesses/zed-agent.md Zed Agent appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Zed and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Continue (desktop) HTML: https://canmyagentuse.com/harnesses/continue Markdown: https://canmyagentuse.com/harnesses/continue.md Continue appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Continue and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Cline (desktop) HTML: https://canmyagentuse.com/harnesses/cline Markdown: https://canmyagentuse.com/harnesses/cline.md Cline appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Cline and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## JetBrains AI (desktop) HTML: https://canmyagentuse.com/harnesses/jetbrains-ai Markdown: https://canmyagentuse.com/harnesses/jetbrains-ai.md JetBrains AI appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of JetBrains and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Warp (desktop) HTML: https://canmyagentuse.com/harnesses/warp Markdown: https://canmyagentuse.com/harnesses/warp.md Warp appears here as a **Desktop** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Warp and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Claude CLI (cli) HTML: https://canmyagentuse.com/harnesses/claude-cli Markdown: https://canmyagentuse.com/harnesses/claude-cli.md Claude CLI appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Anthropic and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## ChatGPT CLI (cli) HTML: https://canmyagentuse.com/harnesses/chatgpt-cli Markdown: https://canmyagentuse.com/harnesses/chatgpt-cli.md ChatGPT CLI appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of OpenAI and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Codex CLI (cli) HTML: https://canmyagentuse.com/harnesses/codex-cli Markdown: https://canmyagentuse.com/harnesses/codex-cli.md Codex CLI appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of OpenAI and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## OpenCode (cli) HTML: https://canmyagentuse.com/harnesses/opencode Markdown: https://canmyagentuse.com/harnesses/opencode.md OpenCode appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of OpenCode and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Gemini CLI (cli) HTML: https://canmyagentuse.com/harnesses/gemini-cli Markdown: https://canmyagentuse.com/harnesses/gemini-cli.md Gemini CLI appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Google and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Aider (cli) HTML: https://canmyagentuse.com/harnesses/aider Markdown: https://canmyagentuse.com/harnesses/aider.md Aider appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Aider and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Goose (cli) HTML: https://canmyagentuse.com/harnesses/goose Markdown: https://canmyagentuse.com/harnesses/goose.md Goose appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Block and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Copilot CLI (cli) HTML: https://canmyagentuse.com/harnesses/copilot-cli Markdown: https://canmyagentuse.com/harnesses/copilot-cli.md Copilot CLI appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Microsoft and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## Amp (cli) HTML: https://canmyagentuse.com/harnesses/amp-cli Markdown: https://canmyagentuse.com/harnesses/amp-cli.md Amp appears here as a **CLI** column so a capability page can show this harness next to others. The product name is a plain-text nominative reference. Can My Agent Use is independent of Sourcegraph and does not claim certification, partnership, or official status. Support cells for this surface are published only when a dated note cites reviewable public evidence; every other cell remains **unknown**. --- ## A2UI HTML: https://canmyagentuse.com/specs/a2ui Markdown: https://canmyagentuse.com/specs/a2ui.md A2UI describes user interfaces as structured streamed data. It is a UI representation layer, distinct from tool protocols, agent-to-agent task exchange, and backend-to-application event protocols. The catalog keeps the stable 0.9 line and the 1.0 candidate explicit. Renderer coverage, component vocabulary, transport, and extension behavior require revision-specific evidence. --- ## AG-UI HTML: https://canmyagentuse.com/specs/ag-ui Markdown: https://canmyagentuse.com/specs/ag-ui.md AG-UI focuses on interaction between an agent backend and an application. MCP can supply tools and context, while A2A can exchange work between agents; those layers can coexist with AG-UI. Compatibility records should identify protocol revision, event families, transport, state synchronization behavior, and application/backend role. --- ## Agent Client Protocol HTML: https://canmyagentuse.com/specs/acp Markdown: https://canmyagentuse.com/specs/acp.md ACP defines a boundary between coding agents and clients such as editors. Client and agent are asymmetric roles; a product can participate in one without implementing the other. Assertions should identify role, transport, protocol revision, distribution method, and target release. A CLI flag or registry listing is not by itself full protocol conformance. --- ## Agent Host Protocol HTML: https://canmyagentuse.com/specs/ahp Markdown: https://canmyagentuse.com/specs/ahp.md AHP addresses durable agent sessions and client access to those sessions. It is distinct from ACP’s editor-to-coding-agent boundary and from A2A task exchange. The research snapshot treats AHP as preview. Compatibility claims should include a precise implementation version and feature flag or preview state until the protocol and host behavior stabilize. --- ## Agent Plugins HTML: https://canmyagentuse.com/specs/agent-plugins Markdown: https://canmyagentuse.com/specs/agent-plugins.md Agent Plugins combines a portable manifest core with optional skills and MCP configuration. Namespaced extensions allow hosts to preserve non-portable behavior without redefining the core format. Open Agent Plugins, Claude plugins, Cursor plugins, Codex plugins, Kiro Powers, and VS Code agent plugins are related packaging families, not automatically interchangeable formats. The catalog keeps their install, trust, hook, and lifecycle semantics distinct. --- ## Agent Skills HTML: https://canmyagentuse.com/specs/agent-skills Markdown: https://canmyagentuse.com/specs/agent-skills.md An Agent Skill is a directory whose `SKILL.md` provides metadata and instructions. Optional references, assets, and scripts can extend the package. The catalog separates discovery from progressive loading, resource resolution, script execution, and scope precedence. Recognizing a skill directory does not establish identical lifecycle or trust behavior across hosts. --- ## Agent2Agent Protocol HTML: https://canmyagentuse.com/specs/a2a Markdown: https://canmyagentuse.com/specs/a2a.md A2A occupies the agent-to-agent layer. It does not replace MCP’s tool/context boundary, ACP’s editor/agent boundary, or a UI event protocol. Catalog records should identify specification revision, role, discovery behavior, task lifecycle, streaming support, and protocol binding rather than collapsing the family into one support check. --- ## AGENTS.md HTML: https://canmyagentuse.com/specs/agents-md Markdown: https://canmyagentuse.com/specs/agents-md.md AGENTS.md is an instruction-file convention, not a runtime protocol. Hosts can differ in where they search, whether nested files override or merge, which filenames coexist, and how instructions interact with user or system rules. The catalog therefore treats root discovery, nested discovery, precedence, and scope as separate capabilities. Recognizing the same filename does not establish identical semantics. --- ## MCP Apps HTML: https://canmyagentuse.com/specs/mcp-apps Markdown: https://canmyagentuse.com/specs/mcp-apps.md MCP Apps defines an interactive UI layer on top of MCP. The host discovers a declared UI resource, renders it in an isolated frame, and mediates communication with the tool and host. Tool calling and app rendering remain separate catalog capabilities. Host listing, plan availability, developer-mode access, sandbox behavior, and extension revision should remain visible in the evidence. --- ## Model Context Protocol HTML: https://canmyagentuse.com/specs/mcp Markdown: https://canmyagentuse.com/specs/mcp.md MCP is not one boolean capability. Tools, resources, prompts, sampling, elicitation, roots, tasks, transports, and authorization are independent questions with direction-specific roles. Catalog assertions should pin this revision (or another named revision), identify the exact harness surface, and retain transport, authorization, plan, policy, and runtime conditions as qualifiers. A broad product statement that it “supports MCP” is discovery evidence, not proof of every primitive. --- ## NLWeb HTML: https://canmyagentuse.com/specs/nlweb Markdown: https://canmyagentuse.com/specs/nlweb.md NLWeb combines semantic web data with natural-language-oriented interfaces and MCP-related integration. It occupies a web exposure layer, not a general replacement for every agent protocol. The research snapshot records an early TypeSpec line. Assertions should name implementation, revision, endpoint behavior, data assumptions, and evidence class. --- ## WebMCP HTML: https://canmyagentuse.com/specs/webmcp Markdown: https://canmyagentuse.com/specs/webmcp.md WebMCP sits at the website-to-agent boundary. It is not the same thing as configuring a remote MCP server, even when product language uses similar “tools” terminology. The research snapshot records a draft Community Group Report. Any positive implementation assertion should name the draft revision, runtime or browser version, flags, origin/security conditions, and evidence class. --- ## About HTML: https://canmyagentuse.com/about Markdown: https://canmyagentuse.com/about.md Can My Agent Use is an independent compatibility catalog for engineers choosing an agent product. It answers one question: **does this exact product support this capability, and which public source supports the status?** The task is familiar: search a capability, scan web, desktop, and CLI columns, read the limits, and open the cited source. Caniuse.com inspired that information structure. The implementation, night-ledger visual system, mark, assets, and copy are original. ## Who it serves People can use the search, matrices, product pages, specification index, evidence ledger, contradiction register, test definitions, and comparison table. Agents can start at `/llms.txt`, fetch a `.md` copy, or query the versioned JSON API. Every format comes from the same content records. Machine-readable copies do not receive different claims. ## What the table promises Unknown is the default. A cell changes only when a dated note cites public evidence for one capability revision, exact product target, and environment profile. Partial means the capability has a documented limit. The sourced-cell percentage measures catalog completeness. It does not measure adoption, popularity, quality, conformance, or market share. The catalog separates workspace instructions, skills, package formats, tool protocols, web exposure, interactive UI, editor protocols, session hosts, and agent-to-agent exchange. A claim about one layer does not prove support for another. A brand-level claim does not prove every child capability. ## Independent by design Names such as ChatGPT, Claude, Cursor, Gemini, Grok, Grok Bot, Copilot, Codex, OpenCode, and OpenWork Desktop appear as plain text to identify products. Can My Agent Use is not affiliated with, endorsed by, certified by, or official to any listed provider. The site does not use provider screenshots, fonts, product shells, or proprietary interface assets. Any secondary provider marks are locally copied, provenance-recorded nominative references with an original monogram fallback; they do not imply affiliation. --- ## Agent harness evaluation guide HTML: https://canmyagentuse.com/harness-guide Markdown: https://canmyagentuse.com/harness-guide.md Choosing an agent harness is not the same as choosing a model. The harness determines which model and context budget are actually available, how files and tools reach it, what work can continue or be delegated, which actions require approval, what data leaves the environment, how usage is billed, and whether the resulting work can be audited or moved elsewhere. Start with the exact **product surface × plan × model or mode × environment × region × organization policy**. Do not combine a provider’s web chat, desktop application, editor extension, CLI, API, and enterprise workspace into one answer. ## Fast decision checklist | Decision area | Minimum useful questions | | --- | --- | | Product boundary | Which exact web, desktop, editor, or CLI surface? Which release or observation date, plan, account type, region, and policy state? | | Models and context | Which model actually runs? Can you choose it? What are the advertised and effective input limits, output limit, context meter, compaction behavior, and fallback policy? | | Caching and cost | Is repeated context cached? Is reuse automatic or controllable? Can you see cache reads, writes, hits, tokens, latency, and price? | | Files and media | Which document, image, audio, and video formats reach the model? What is extracted or sampled? What are the byte, file-count, page, duration, and storage limits? | | Tools and environment | Can it read and write workspace files, run a terminal, use a browser, call connectors or MCP, access the network, and isolate side effects? | | Delegated execution | Can it create child agents? Can they run concurrently or recursively? What context, models, tools, files, secrets, and approvals do they inherit? | | Security and data | What is retained, used for improvement, regionally processed, encrypted, logged, and deletable? Can admins enforce policy and scope secrets? | | Operations | What is included in the plan? Which quotas, rate limits, timeouts, retries, checkpoints, resumability, notifications, and status surfaces exist? | | Collaboration and exit | Can teams share and assign work with roles? Can conversations, provenance, files, patches, and structured artifacts be exported? | | Evidence quality | Is every claim exact-surface, dated, public, and scoped? Is it documented, tested, observed, inferred, or merely not found? | ## 1. Establish the compatibility target Record the product name and exact surface first. For installed software, record the release, operating system, architecture, extension or editor version, and relevant configuration. For continuously deployed web products, use a dated hosted observation. Include account type, subscription, region, organization policy, feature flags, model selection, and local versus hosted execution. This prevents a common category error: copying an API or model capability into a chat, desktop, editor, or CLI column without evidence that the harness exposes it. ## 2. Separate model access from context behavior Ask which models can be selected, which are automatically routed, whether reasoning effort can be controlled, and whether fallback can silently change the model. Then record both advertised and effective context limits. An effective budget may be smaller because the harness reserves capacity for output, hidden instructions, tool schemas, memory, retrieval, attachments, images, video frames, or earlier messages. Record the unit, model and mode, input cap, output cap, observation date, and boundary behavior. Acceptance of a large prompt does not prove that all of it remains available; test rejection, truncation, omission, retrieval, and automatic compaction separately. For long sessions, determine when compaction occurs, whether it is visible, what it preserves, whether it can be triggered or configured, and how commitments survive the boundary. A generic model-card context number does not answer those questions. ## 3. Treat prompt caching as three claims First ask whether eligible repeated prompt or context prefixes receive documented reuse. Then ask whether the harness exposes controls such as breakpoints, cache objects, retention choices, refresh, or invalidation. Finally ask whether an operator can verify reads, writes, hits, misses, cached tokens, latency, and cost. Do not confuse model prompt caching with browser or HTTP caches, downloaded-file caches, vector indexes, retrieval caches, build caches, or a chat merely retaining its history. Record minimum prefix size, exact-match rules, supported models, lifetime, isolation boundary, invalidation, read and write pricing, and whether the harness preserves a cache-friendly prompt order. ## 4. Follow every file into model context A file picker proves upload, not understanding. For each document, image, audio, or video format, identify whether the original reaches a multimodal model, text is extracted, pages or frames are rendered, audio is transcribed, a tool receives the file, or only metadata is available. Record accepted extensions and MIME types; per-file and aggregate bytes; files per message or project; pages, duration, resolution, and storage; preprocessing; truncation or sampling; model and plan restrictions; latency; retention; and failure behavior. For video, test temporal order, scene changes, audio-visual alignment, on-screen text, and time-specific citations. A correct transcript answer does not prove visual or temporal understanding. Workspace access is different from upload. A coding agent may read repository files without accepting an interactive attachment, while a hosted chat may accept files without exposing a filesystem. ## 5. Map tools, side effects, and approval boundaries Inventory filesystem search and writes, patches, terminal execution, code sandboxes, git and pull requests, browser automation, computer use, web fetch, connectors, MCP roles and transports, plugins, skills, and structured output. For each tool path, record authentication, network policy, filesystem scope, sandbox lifetime, environment variables, secrets, approval rules, allowed destinations, logs, failure behavior, and whether an action is proposed or verified after execution. Read access does not imply write access; browser navigation does not imply general desktop control; MCP tools do not imply MCP prompts, resources, sampling, Apps, or every transport. ## 6. Decompose subagents and background work Basic delegation establishes only one child execution boundary. Evaluate nesting, concurrent children, queue and fan-out limits, context inheritance, isolation, per-child model choice, MCP and filesystem permissions, secret access, approval ownership, lifecycle controls, and result handoff independently. Ask whether the operator can list, inspect, pause, resume, cancel, retry, and time-limit children; whether cancellation reaches descendant tools; how simultaneous writers avoid conflicts; and whether results return as traceable files, diffs, citations, status, and errors rather than unstructured prose. For background or scheduled work, record where it executes after the client closes, maximum duration, checkpoints, reconnect and resume behavior, notifications, plan requirements, and what happens to pending approvals. ## 7. Trace data and administrative control end to end Follow prompts, uploads, repository content, retrieved data, tool results, memory, generated artifacts, feedback, telemetry, logs, caches, safety copies, and backups. For each, ask about retention, deletion, training or improvement use, human review, storage and processing regions, encryption, customer-managed keys, support access, and subprocessors. Then evaluate managed secrets, identity, role and policy enforcement, audit logs, connector authorization, local-only or offline behavior, and organization controls over models, tools, sharing, uploads, memory, data use, and sub-agent fan-out. A native app is not automatically local-only, and a provider-wide certification is not a substitute for exact-surface data-flow evidence. ## 8. Price the real workload and its failure modes Record plan price and billing period, included usage, model premiums, tool or connector charges, hosted compute, storage, media processing, sub-agent usage, overages, hard caps, seat minimums, and enterprise-only terms. Pair that with per-run and aggregate meters so estimates can be reconciled with actual use. Document request, token, concurrency, upload, tool, and rolling quotas; foreground and background timeouts; approval and idle deadlines; retry and backoff; model fallback; checkpoints; idempotency; rollback or reconciliation; and preservation of partial results. A safe retry policy must distinguish reads from externally visible writes such as sending a message, changing a database, deploying code, or purchasing an item. ## 9. Check collaboration and the exit path Separate a public share link from named access, comments, edits, continuation rights, shared project context, granular roles, and accountable handoff. Record revocation, expiration, external access, recipient account requirements, inherited connector permissions, ownership transfer, and audit attribution. Test export before committing important work. A useful run export contains roles, timestamps, models, instructions, tool calls and results, approvals, citations, attachments, sub-agent relationships, errors, usage, and status. Artifact export should preserve original files, patches, repositories, reports, media, structured results, metadata, and checksums in a form that remains usable outside the harness. ## 10. Keep evidence proportional to the claim Prefer current public product documentation, release notes, maintained first-party repositories, standards text, and reproducible tests. Record what was reviewed and when. Search snippets, rumors, private beta screens, provider-family assumptions, and silence do not justify a definitive support state. Use **unknown** when the exact claim has not passed the evidence bar. Use **partial** when the capability is constrained by plan, platform, model, region, rollout, policy, runtime, transport, format, size, or another material boundary. Reserve **no** for explicit current evidence of non-support. Keep lifecycle labels such as planned, preview, stable, and deprecated separate from compatibility. ## Minimum decision record Before selecting a harness, preserve: 1. Exact surface, release or observation date, plan, region, policy state, environment, model, and mode. 2. Required capabilities and hard limits, including context, output, uploads, concurrency, duration, quotas, and cost ceiling. 3. Data classes and permitted storage, processing, retention, training, sharing, and export paths. 4. Required tool, connector, secret, approval, audit, and administrative controls. 5. Failure, recovery, checkpoint, notification, and human-handoff expectations. 6. Dated sources, tested observations, unresolved contradictions, and every remaining unknown. That record turns a vague product preference into a reviewable engineering and procurement decision. --- ## Methodology HTML: https://canmyagentuse.com/methodology Markdown: https://canmyagentuse.com/methodology.md A support assertion answers one narrow question for a **capability revision × exact product target × environment profile**. A rendered cell may be `yes`, `partial`, `no`, `unknown`, or `na`. - `yes` means current public documentation demonstrates the row’s capability without a material limit. - `partial` means the capability exists with a documented plan, platform, transport, rollout, environment, or interaction limit. - `no` requires an explicit current source saying the capability is unsupported. Silence is not evidence of no. - `unknown` is the default whenever evidence has not been reviewed or remains ambiguous. - `na` means the row does not apply to that kind of harness and still needs public evidence. ## Capability families and atomic rows Broad labels such as MCP, plugins, Agent Skills, subagents, model and context management, file inputs, data security, operational reliability, collaboration, persistent instructions, and WebMCP are capability families, not compatibility cells. A family has independently testable atomic children. For example, MCP tool calling does not imply MCP prompts, resources, sampling, authorization, transport, or Apps support. Likewise, basic subagent delegation does not imply concurrency, nesting, inherited tools, isolated context, independent model choice, or structured handoff. Family progress is derived from child rows for each exact product. The catalog reports both reviewed children and supported-or-partial children, plus the full status breakdown. It never authors or infers one umbrella family status. Atomic rows may also declare typed `requires`, `extends`, `alternative`, or `related` edges. Siblings are derived from their shared family parent. ## Evidence contract Every non-unknown cell is authored as a version entry with one or more note IDs, an explicit target, an environment profile, structured qualifiers, and typed evidence references. Each evidence reference resolves to a stable resource ID and records when it was reviewed. The resource list links to public vendor documentation, release notes, maintained first-party repositories, or the relevant open-standard documentation. Targets are versioned releases when a stable release is public. Continuously deployed web products use a dated hosted observation. When only current local-product documentation is available, the record says `dated-documentation` rather than pretending a particular binary was reproduced. Evidence classes stay separate: **documented**, **vendor-attested**, **listed**, **tested**, **reported**, **inferred**, and **not found** do not carry the same authority. Catalog v1 publishes documentation and listing evidence. The public test registry contains proposed definitions only; Can My Agent Use has not executed runtime conformance tests against these harnesses. Plan, policy, region, authorization, transport, protocol revision, role, feature flag, runtime, preview, and vendor-extension limits remain structured qualifiers. They are not folded into prose and lost when the matrix is rendered. Compatibility status and product lifecycle are separate. An exact cell can be currently unsupported while its lifecycle is `planned`, or partially supported while it is `experimental` or `preview`. Lifecycle values are `untracked`, `requested`, `planned`, `experimental`, `preview`, `stable`, and `deprecated`; they require the same evidence discipline as the associated compatibility assertion. ## Limits and measured values Some of the most useful harness facts are numeric: advertised and effective context tokens, maximum output, file bytes and count, document pages, audio or video duration, concurrent subagents, requests or tokens per reset window, run duration, storage, retention, and price. A number is not portable without its unit and scope. Every measured value should identify the exact harness, model or mode, plan, environment, region or policy when relevant, what the value limits, whether it is a documented maximum or a tested observation, the reset or billing period, the review date, and the boundary behavior. A model API limit is not copied into a chat, desktop, editor, or CLI surface without product evidence. When an advertised maximum and an effective usable limit differ, preserve both and explain the reservation, truncation, sampling, or compaction behavior. Prompt caching is split into three claims: reuse, operator controls, and telemetry. File upload is split from model-visible understanding and from preprocessing disclosure. These distinctions keep a paperclip icon, a provider model card, or a generic caching statement from becoming a broader compatibility claim than the evidence supports. Do not source a claim from rumors, search snippets without a stable page, community recollection, private beta screens, or screenshots of interfaces that are not public. When a document describes a model or API rather than the named harness, it does not automatically prove harness support. Community reports can surface an investigation or contradiction, but do not change a definitive support state without authoritative or reproducible corroboration. When sources conflict, the contradiction stays public instead of being smoothed into a color. ## Tracks and coverage Tracks (`current`, `preview`) are catalog version stacks. They are not a claim that a vendor publishes channels with those names, and evidence for current never fills preview automatically. The feature-page percentage is the share of published product tracks that have moved off unknown. It measures catalog coverage. It does not measure usage, adoption, quality, or market share. The first review covers MCP tools, MCP Apps, workspace files, terminal access, computer use, and skills for a small set of products. All other cells stay unknown until reviewed. ## Corrections Public documentation changes. A correction should supersede the affected statement, update the review date, preserve a clear explanation of the documented limit, and replace stale resource links without erasing the reason for the change. Contributors should follow `CONTRIBUTING.md`; the shared schema rejects non-unknown cells without targets, environment profiles, notes, and typed evidence. The product deliberately does not publish a synthetic harness score or fabricated market-share figure. Useful operational metrics are evidence freshness, sourced-cell coverage, contradiction age, revision pinning, source diversity, review latency, and future conformance-test coverage. --- ## Privacy HTML: https://canmyagentuse.com/privacy Markdown: https://canmyagentuse.com/privacy.md Can My Agent Use does not ask you to create an account and does not collect votes, support reports, search queries, or form submissions. Search and comparison interactions run in your browser. The site does not currently use first-party behavioral analytics, advertising trackers, or marketing pixels. The public catalog is hosted on Cloudflare Workers. Like other hosting providers, Cloudflare may process ordinary request metadata needed to deliver and protect the service, such as IP address, user agent, requested URL, timing, and security signals, under Cloudflare's applicable terms and privacy documentation. Outbound evidence links take you to third-party sites. Their privacy practices apply after you leave this catalog. If collection is added later, this page and the product UI must be updated before launch to explain the data, purpose, retention, and controls. --- ## Catalog foundation is up HTML: https://canmyagentuse.com/news/catalog-foundation Markdown: https://canmyagentuse.com/news/catalog-foundation.md The site now supports the main task: search a capability, scan product columns, and read the notes. Product names identify catalog columns. They do not imply affiliation. Support stays unknown until sourced. --- ## Evidence Atlas v1 is ready HTML: https://canmyagentuse.com/news/evidence-atlas-v1 Markdown: https://canmyagentuse.com/news/evidence-atlas-v1.md The first reviewed wave covers MCP tools, MCP Apps, workspace files, terminal commands, computer use, and Agent Skills across selected exact harness surfaces. Every non-unknown statement records a target, environment profile, structured conditions, dated notes, and typed public evidence. The site now exposes twelve revision-aware specification references, an evidence ledger, a proposed-but-not-run conformance test registry, an ecosystem change feed, and a contradiction register. HTML, Markdown, JSON, OpenAPI, sitemap, and `llms.txt` surfaces are generated from the same catalog records. Unknown remains the default. Documentation evidence remains distinct from runtime testing, and the product publishes no synthetic compatibility score or fabricated usage share. --- ## Web, desktop, and CLI columns HTML: https://canmyagentuse.com/news/harness-columns Markdown: https://canmyagentuse.com/news/harness-columns.md Columns are grouped by web, desktop, and CLI product type, not by vendor. Each column publishes the tracks declared for that exact product surface. Unreviewed tracks and cells render as unknown rather than inheriting support from another surface. Listed first-party product families include ChatGPT, Claude, Gemini, Copilot, Cursor, OpenCode, **OpenWork Desktop**, Windsurf, Zed, Continue, Cline, Aider, Goose, Warp, Amp, Devin, and Replit Agent. “First-party” identifies a publisher's own product surface; it does not imply that Can My Agent Use is affiliated with or endorsed by that publisher. --- ## Unknown is the published default HTML: https://canmyagentuse.com/news/unsourced-cells Markdown: https://canmyagentuse.com/news/unsourced-cells.md The feature page shows how many version tracks have a status other than unknown. That number is a catalog-completeness signal. It is not a global usage share and not a product score. --- --- title: "Evidence Atlas" canonical: "https://canmyagentuse.com/atlas" contentKind: "page" locale: "en" description: "The complete 2026-08-28 documentation-backed compatibility seed, kept separate from the curated web/desktop/CLI matrix so every claim retains its exact surface." llmSummary: "The complete 2026-08-28 documentation-backed compatibility seed, kept separate from the curated web/desktop/CLI matrix so every claim retains its exact surface." publishedAt: null updatedAt: "2026-08-28T00:00:00.000Z" verifiedAt: "2026-08-28" tags: ["evidence-atlas", "research"] --- # Evidence Atlas The complete 2026-08-28 documentation-backed compatibility seed, kept separate from the curated web/desktop/CLI matrix so every claim retains its exact surface. Research cutoff: 2026-08-28. Documentation and registry evidence are not runtime certification. Unknown means the package did not establish a defensible claim; it does not mean unsupported. ## Snapshot - Vendors: 51 - Products: 60 - Exact surfaces: 93 - Atomic features: 49 - Compatibility cells: 4557 - Supported: 790 - Partial: 30 - Explicit unsupported: 42 - Unknown: 3695 - Source records: 183 - Long-form assertions: 959 ## Vendor coverage | Vendor | Tier | Surfaces | Supported | Partial | Unsupported | Unknown | | --- | --- | ---: | ---: | ---: | ---: | ---: | | [Agoragentic](/atlas/vendors/agoragentic) | C | 1 | 2 | 0 | 0 | 47 | | [Aider](/atlas/vendors/aider) | D | 1 | 2 | 1 | 0 | 46 | | [Alibaba / Qwen](/atlas/vendors/alibaba-qwen) | A | 1 | 15 | 0 | 0 | 34 | | [Amp](/atlas/vendors/amp) | C | 1 | 2 | 0 | 0 | 47 | | [Anthropic](/atlas/vendors/anthropic) | A | 2 | 24 | 3 | 0 | 71 | | [Augment Code](/atlas/vendors/augment) | A | 2 | 28 | 0 | 0 | 70 | | [Autohand](/atlas/vendors/autohand) | C | 1 | 2 | 0 | 0 | 47 | | [AWS / Kiro](/atlas/vendors/aws-kiro) | A | 4 | 62 | 2 | 5 | 127 | | [Block / goose](/atlas/vendors/block-goose) | A | 3 | 31 | 2 | 0 | 114 | | [Cline](/atlas/vendors/cline) | A | 2 | 26 | 2 | 4 | 66 | | [Cognition](/atlas/vendors/cognition) | A | 4 | 34 | 2 | 0 | 160 | | [Compass AI / Nova](/atlas/vendors/compass-ai) | C | 1 | 2 | 0 | 0 | 47 | | [Continue](/atlas/vendors/continue) | A | 3 | 9 | 0 | 10 | 128 | | [Corust](/atlas/vendors/corust) | C | 1 | 2 | 0 | 0 | 47 | | [crow-cli](/atlas/vendors/crow) | C | 1 | 2 | 0 | 0 | 47 | | [Cursor](/atlas/vendors/cursor) | A | 3 | 55 | 2 | 0 | 90 | | [DimCode](/atlas/vendors/dimcode) | C | 1 | 2 | 0 | 0 | 47 | | [Dirac](/atlas/vendors/dirac) | C | 1 | 2 | 0 | 0 | 47 | | [Factory AI](/atlas/vendors/factory) | A | 2 | 19 | 0 | 1 | 78 | | [fast-agent](/atlas/vendors/fast-agent) | C | 1 | 2 | 0 | 0 | 47 | | [GitHub](/atlas/vendors/github) | A | 4 | 54 | 0 | 3 | 139 | | [Google](/atlas/vendors/google) | A | 2 | 15 | 2 | 0 | 81 | | [Harn](/atlas/vendors/harn) | C | 1 | 2 | 0 | 0 | 47 | | [JetBrains](/atlas/vendors/jetbrains) | A | 3 | 13 | 1 | 0 | 133 | | [Kilo](/atlas/vendors/kilo) | A | 2 | 15 | 0 | 0 | 83 | | [LangChain / DeepAgents](/atlas/vendors/deepagents) | C | 1 | 2 | 0 | 0 | 47 | | [Microsoft / Visual Studio Code](/atlas/vendors/microsoft-vscode) | A | 1 | 14 | 0 | 0 | 35 | | [Minion Code](/atlas/vendors/minion) | C | 1 | 2 | 0 | 0 | 47 | | [Mistral AI](/atlas/vendors/mistral) | A | 1 | 10 | 0 | 1 | 38 | | [Moonshot AI](/atlas/vendors/moonshot) | A | 2 | 22 | 0 | 1 | 75 | | [NanoClaw](/atlas/vendors/nanoclaw) | B | 1 | 9 | 0 | 0 | 40 | | [Nous Research / Hermes](/atlas/vendors/nous-hermes) | A | 2 | 17 | 0 | 0 | 81 | | [OpenAI](/atlas/vendors/openai) | A | 6 | 63 | 1 | 5 | 225 | | [OpenClaw](/atlas/vendors/openclaw) | A | 3 | 20 | 1 | 0 | 126 | | [OpenCode](/atlas/vendors/opencode) | A | 3 | 54 | 0 | 0 | 93 | | [OpenHands](/atlas/vendors/openhands) | A | 4 | 34 | 5 | 0 | 157 | | [OpenWork](/atlas/vendors/openwork) | A | 1 | 18 | 4 | 3 | 24 | | [pi](/atlas/vendors/pi) | C | 1 | 2 | 0 | 0 | 47 | | [Poolside](/atlas/vendors/poolside) | C | 1 | 2 | 0 | 0 | 47 | | [Qoder](/atlas/vendors/qoder) | A | 1 | 16 | 0 | 0 | 33 | | [Replit](/atlas/vendors/replit) | A | 2 | 12 | 0 | 0 | 86 | | [Roo Code](/atlas/vendors/roo-code) | B | 1 | 0 | 0 | 6 | 43 | | [siGit Code](/atlas/vendors/sigit) | C | 1 | 2 | 0 | 0 | 47 | | [Snowflake](/atlas/vendors/snowflake) | C | 1 | 2 | 0 | 0 | 47 | | [Stakpak](/atlas/vendors/stakpak) | C | 1 | 2 | 0 | 0 | 47 | | [Tencent Cloud / CodeBuddy](/atlas/vendors/codebuddy) | C | 1 | 2 | 0 | 0 | 47 | | [VT Code](/atlas/vendors/vt-code) | C | 1 | 2 | 0 | 0 | 47 | | [Warp](/atlas/vendors/warp) | A | 3 | 25 | 1 | 0 | 121 | | [xAI](/atlas/vendors/xai) | A | 2 | 25 | 0 | 0 | 73 | | [Z.AI / GLM Agent](/atlas/vendors/zai) | C | 1 | 2 | 0 | 0 | 47 | | [Zed Industries](/atlas/vendors/zed) | A | 2 | 9 | 1 | 3 | 85 | ## Machine-readable entry points - [Atlas JSON](/api/v1/atlas.json) - [Source ledger](/atlas/sources) - [Curated matrix](/matrix)