Capabilities
Each capability has an HTML article, a Markdown twin, and a JSON record. Support cells stay unknown until sourced.
collaboration
Accountable human handoffTransfer responsibility to a named person or queue with context, status, deadline, and acknowledgement.
security-privacy
Administrative policy controlsCentrally enforce model, tool, connector, data, sharing, approval, and autonomy boundaries.
interfaces
Agent Plugins failure isolationReject invalid components narrowly without disabling valid components in the same package.
interfaces
Agent Plugins packaged MCPLoad MCP server configuration contained in an Agent Plugins package.
interfaces
Agent Plugins packaged skillsLoad portable Agent Skills contained in an Agent Plugins package.
interfaces
Agent Plugins portable manifestLoad the published portable plugin.json manifest and its core components.
runtime
Agent SkillsTrack every important Agent Skills behavior independently.
runtime
Agent Skills core and discoveryRecognize, discover, and load reusable instruction packages centered on SKILL.md.
runtime
Agent Skills manual invocationLet a user or agent explicitly select and activate a named skill.
runtime
Agent Skills progressive loadingLoad skill metadata first and defer full instructions or resources until relevant.
runtime
Agent Skills references and assetsResolve supporting reference files, templates, and assets from a skill package.
runtime
Agent Skills scriptsResolve and execute optional scripts included with an Agent Skill under explicit runtime policy.
interfaces
AGENTS.md root instructionsDiscover and apply the cross-agent AGENTS.md convention at repository scope.
collaboration
Artifact and workspace exportTransfer generated files, patches, repositories, reports, media, and structured results with provenance.
perception
Audio file upload and understandingUpload recorded audio and expose its supported speech, timing, speaker, or acoustic content.
security-privacy
Audit logs and actor attributionAttribute and export security-relevant user, agent, tool, data, approval, and policy events.
models-context
Automatic context compactionContinue long sessions by automatically summarizing or pruning older context.
runtime
Background agentsKeep a run going after the operator leaves the session.
tools
Browser automationDrive a browser for clicks, forms, and navigation.
tools
Code executionRun generated code in a sandbox.
collaboration
Collaboration and portabilityTrack team access, shared context, accountable handoff, and portable outputs independently.
runtime
Computer useInspect and act on a visual interface through screenshots and input actions.
runtime
Concurrent subagentsRun multiple child-agent tasks concurrently with explicit queue and fan-out limits.
tools
ConnectorsReach mail, chat, calendar, or issue trackers the user authorizes.
models-context
Context compaction controlsTrigger, inspect, configure, or disable context compaction.
models-context
Context usage visibilityInspect consumed or remaining context budget during a conversation or run.
models-context
Context window and effective input limitRecord the advertised and effective maximum input budget by model and harness mode.
collaboration
Controlled conversation sharingShare a run with explicit identity, permission, revocation, redaction, and update semantics.
collaboration
Conversation and run exportExport a documented, portable run record with messages, actions, provenance, and status.
interfaces
Custom instructionsPersist project or user instruction files.
security-privacy
Data residency and processing location controlsControl and document storage, inference, transient processing, logging, and backup regions.
security-privacy
Data retention and deletion controlsConfigure and verify retention and deletion across user content, logs, memory, and backups.
security-privacy
Data security and administrative controlsTrack data handling and administrative control boundaries as independent capabilities.
tools
Diffs and patchesPropose and apply file diffs without a full git flow.
security-privacy
Encryption and key controlsRecord transport and at-rest encryption plus customer key scope, rotation, and revocation.
operations
Failure recovery and safe retryRecover from model, tool, connector, worker, and network failures without unsafe duplicate effects.
perception
File and media inputsTrack accepted file types, model-visible content, preprocessing, and upload limits separately.
perception
File upload limits and quota visibilityRecord numeric attachment size, count, page, duration, storage, and rolling quotas.
tools
Git and pull requestsInspect diffs and open or review pull requests.
runtime
Human approvalPause before a tool or write until a person confirms.
perception
Image upload and pasteAttach, paste, or select an image whose visual content reaches the model.
models-context
Local modelsRun against a model hosted on the operator's machine.
security-privacy
Local-only and offline operationRun a documented workflow without required hosted inference, storage, control-plane, or telemetry dependencies.
runtime
Long-term memoryPersist notes across sessions.
interfaces
MCP AppsRender an interactive interface returned by an MCP server inside the product.
interfaces
MCP cancellationCancel an in-flight MCP request and handle cancellation correctly.
interfaces
MCP client roleConsume capabilities exposed by MCP servers.
interfaces
MCP elicitationHandle structured user-input requests initiated by an MCP server.
interfaces
MCP legacy HTTP and SSEConnect through the superseded HTTP plus SSE transport when migration compatibility is required.
interfaces
MCP list-changed notificationsRefresh capability lists when a server announces dynamic changes.
interfaces
MCP loggingReceive structured log messages emitted by an MCP server.
interfaces
MCP OAuth authorizationComplete the MCP authorization flow for protected remote servers.
interfaces
MCP progress notificationsReceive progress updates for long-running requests.
interfaces
MCP prompt-list notificationsRefresh prompt templates after a server announces changes.
interfaces
MCP promptsInvoke reusable MCP prompt templates.
interfaces
MCP Registry metadataDiscover MCP servers and metadata through the official registry model.
interfaces
MCP resource subscriptionsSubscribe to resource changes and receive update notifications.
interfaces
MCP resourcesRead MCP resources as live context.
interfaces
MCP revision compatibilityImplement or negotiate the exact dated MCP revision.
interfaces
MCP rootsProvide filesystem or workspace roots to an MCP server.
interfaces
MCP samplingHandle server requests for model sampling through the client.
interfaces
MCP server instructionsConsume server-provided instructions during initialization.
interfaces
MCP server roleExpose capabilities from the harness as an MCP server.
interfaces
MCP stdio transportConnect to a local MCP server over standard input and output.
interfaces
MCP Streamable HTTPConnect to an MCP server over the current Streamable HTTP transport.
interfaces
MCP task listingList and page through tasks exposed by the task utility.
interfaces
MCP tasksCreate, inspect, and manage long-running MCP tasks.
interfaces
MCP toolsInvoke operations exposed by a Model Context Protocol server.
interfaces
Model Context ProtocolTrack tools, prompts, resources, and interactive apps as independent parts of MCP.
models-context
Model fallback visibility and controlDetect, configure, or prevent fallback to another model during a run.
models-context
Model selectionChoose the model or documented model class used for a run.
security-privacy
Model training and improvement controlsDisclose and control training, improvement, evaluation, safety review, and feedback use.
models-context
Models and contextTrack model access and the full lifecycle of input, output, compaction, and cached context.
tools
Multi-file editChange several files in one reviewed step.
interfaces
Native plugin systemInstall and load a product-specific extension package.
interfaces
Nested AGENTS.md instructionsApply nested AGENTS.md files with directory-sensitive precedence.
runtime
Nested subagentsAllow a child agent to delegate further work to another child agent.
models-context
Output token limitRecord the maximum generated response budget by exact model and harness mode.
perception
PDF documentsUpload a PDF and make its text, layout, and supported visual content available as context.
runtime
Per-subagent model selectionChoose or constrain each child agent's model independently of the parent.
interfaces
Persistent instructionsTrack portable and product-native instruction discovery independently.
operations
Plan and pricing disclosurePublish plan gates, included usage, variable charges, overages, and relevant billing terms.
interfaces
Plugin marketplace and installationDiscover, install, update, and remove extensions through a registry or marketplace.
interfaces
Plugin systemsTrack portable and product-native extension capabilities independently.
interfaces
Plugin vendor extensionsPreserve namespaced host-specific behavior without representing it as portable core support.
interfaces
Product-native rulesDiscover persistent or path-scoped instructions in the product's own rule format.
runtime
Project-scoped Agent SkillsDiscover skills installed for a repository or workspace.
models-context
Prompt and context cache reuseReuse eligible repeated prompt context with documented latency or billing semantics.
models-context
Prompt cache controlsMark, retain, refresh, or invalidate cached prompt context.
models-context
Prompt cache telemetryInspect cache hits, misses, token counts, latency, or billing effects.
operations
Rate limit and quota disclosureRecord numeric request, token, run, concurrency, upload, and tool quotas with reset behavior.
perception
Realtime voiceSpeak and listen over a live audio session.
models-context
Reasoning effort controlsChoose a documented reasoning depth, effort, or budget for a run.
operations
Resumable runsReconnect or continue after closure, disconnection, restart, pause, or worker migration.
collaboration
Role-based collaboration accessEnforce granular view, edit, run, approve, share, export, billing, and administrative rights.
operations
Run timeout and duration limitsRecord foreground, background, tool, sandbox, idle, approval, and child-agent time limits.
operations
Run, approval, and failure notificationsNotify operators about completion, failure, approvals, budgets, deadlines, and stalled work.
runtime
Sandbox networkAllow or deny outbound network from a tool sandbox.
runtime
Scheduled runsStart a task on a timer without a human prompt.
perception
ScreenshotsCapture the current screen or window as context.
security-privacy
Secrets management and redactionStore and inject credentials with scoping, masking, rotation, approval, and audit controls.
collaboration
Shared projects and team contextShare durable project instructions, files, memory, conversations, agents, and access rules.
interfaces
Streaming outputShow tokens and tool events as they arrive.
interfaces
Structured outputReturn JSON or schema-validated objects.
runtime
Structured subagent result handoffReturn structured child status, findings, citations, files, diffs, and errors to the parent.
runtime
Subagent approval and permission boundariesControl inherited grants, approval prompts, least privilege, escalation, and descendant authority.
runtime
Subagent context isolation and inheritanceDefine and control the conversation, instructions, files, memory, tools, and secrets inherited by a child.
runtime
Subagent delegationDelegate a bounded task to another agent process.
runtime
Subagent lifecycle controlsList, inspect, pause, resume, cancel, retry, and time-limit child-agent tasks.
runtime
Subagent MCP accessAllow a child agent to use MCP tools or declare its own MCP servers.
runtime
Subagent write accessAllow a child agent to create or modify project files.
runtime
SubagentsTrack child-agent execution, isolation, permissions, supervision, and handoff independently.
tools
Terminal commandsPropose or execute shell commands in a project environment.
perception
Text and office document inputUpload common text, word-processing, presentation, and spreadsheet formats as model context.
operations
Usage and cost meteringInspect per-run and aggregate model, cache, tool, compute, storage, and delegated usage.
operations
Usage, cost, and run reliabilityTrack the economic and operational envelope around each harness run.
runtime
User-scoped Agent SkillsDiscover personal or global skills outside the current project.
perception
Video preprocessing disclosureExplain frame, audio, transcript, resizing, truncation, and context-accounting behavior.
perception
Video upload and understandingUpload a video and expose documented visual, audio, transcript, and timing information to the model.
tools
Web fetchRetrieve live web pages or APIs during a run.
interfaces
WebMCPExpose website functionality as structured tools through a browser-mediated agent boundary.
interfaces
WebMCP declarative form toolsExpose semantic HTML forms as structured tools through WebMCP annotations.
interfaces
WebMCP execution cancellationCancel pending tool registration or execution through abort signals and lifecycle rules.
interfaces
WebMCP imperative toolsRegister JavaScript-backed tools through document.modelContext.registerTool().
interfaces
WebMCP permissions policyGate tool exposure across documents and origins through the tools permissions policy.
interfaces
WebMCP security annotationsCommunicate read-only and untrusted-content hints across the site-to-agent boundary.
interfaces
WebMCP tool discoveryExpose the active page tool list and schemas to a WebMCP-aware agent.
interfaces
WebMCP tool executionInvoke a page tool with structured arguments and return its result to the agent.
tools
Workspace filesRead and edit files inside a user-selected local project.
tools
Workspace searchSearch a codebase or knowledge tree before editing.