OpenWork Desktop

Reject invalid components narrowly without disabling valid components in the same package.

Load MCP server configuration contained in an Agent Plugins package.

Load portable Agent Skills contained in an Agent Plugins package.

Load the published portable plugin.json manifest and its core components.

Recognize, discover, and load reusable instruction packages centered on SKILL.md.

Let a user or agent explicitly select and activate a named skill.

Load skill metadata first and defer full instructions or resources until relevant.

Resolve supporting reference files, templates, and assets from a skill package.

Resolve and execute optional scripts included with an Agent Skill under explicit runtime policy.

Discover and apply the cross-agent AGENTS.md convention at repository scope.

Download generated files and other documented outputs.

Upload recorded audio for use as model input.

Record security-relevant activity in an administrative audit log.

Continue long sessions by automatically summarizing or pruning older context.

Keep a run going after the operator leaves the session.

Control a browser for navigation, clicks, forms, and page reading.

Run generated code in a documented execution environment.

Inspect and act on a visual interface through screenshots and input actions.

Run multiple child-agent tasks concurrently with explicit queue and fan-out limits.

Connect to external services authorized by the user or organization.

Trigger, inspect, configure, or disable context compaction.

Inspect consumed or remaining context budget during a conversation or run.

Record the documented context-window limit by product, model, and mode.

Export conversation history in a documented format.

Share a conversation through a documented link or invitation.

Persist instructions at user or project scope.

Select or document geographic storage or processing regions.

Configure or document retention and deletion for covered product data.

Propose or apply file changes as diffs or patches.

Upload common document formats for use as model input.

Record documented encryption key ownership, scope, rotation, and revocation.

Record numeric attachment size, count, page, duration, storage, and rolling quotas.

Pause before a tool action or file change until a person confirms.

Attach, paste, or select an image for use as model input.

Run against a model hosted on the operator's machine.

Persist notes across sessions.

Render an interactive interface returned by an MCP server inside the product.

Cancel an in-flight MCP request and handle cancellation correctly.

Consume capabilities exposed by MCP servers.

Handle structured user-input requests initiated by an MCP server.

Connect through the superseded HTTP plus SSE transport when migration compatibility is required.

Refresh capability lists when a server announces dynamic changes.

Receive structured log messages emitted by an MCP server.

Complete the MCP authorization flow for protected remote servers.

Receive progress updates for long-running requests.

Refresh prompt templates after a server announces changes.

Invoke reusable MCP prompt templates.

Discover MCP servers and metadata through the official registry model.

Subscribe to resource changes and receive update notifications.

Read MCP resources as live context.

Implement or negotiate the exact dated MCP revision.

Provide filesystem or workspace roots to an MCP server.

Handle server requests for model sampling through the client.

Consume server-provided instructions during initialization.

Expose capabilities from the harness as an MCP server.

Connect to a local MCP server over standard input and output.

Connect to an MCP server over the current Streamable HTTP transport.

List and page through tasks exposed by the task utility.

Create, inspect, and manage long-running MCP tasks.

Invoke operations exposed by a Model Context Protocol server.

Switch to another model after a documented failure condition.

Choose the model or documented model class used for a run.

Edit multiple files during one task.

Install and load a product-specific extension package.

Apply nested AGENTS.md files with directory-sensitive precedence.

Allow a child agent to delegate further work to another child agent.

Run a documented workflow without a required network connection.

Centrally configure and enforce product settings for an organization.

Record the maximum generated response budget by exact model and harness mode.

Upload a PDF for use as model input.

Choose or constrain each child agent's model independently of the parent.

Compare documented plan, seat, included-usage, credit, and overage pricing by exact product.

Discover and install extensions through a documented registry or marketplace.

Preserve namespaced host-specific behavior without representing it as portable core support.

Load persistent or path-scoped instructions from a documented product rule format.

Discover skills installed for a repository or workspace.

Mark, retain, refresh, or invalidate cached prompt context.

Inspect cache hits, misses, token counts, latency, or billing effects.

Reuse eligible repeated prompt content under documented caching behavior.

Open or review pull requests through a documented integration.

Record documented product usage limits and reset periods.

Speak and listen over a live audio session.

Choose a documented reasoning depth, effort, or budget for a run.

Assign and enforce permissions through documented roles.

Allow, deny, or restrict outbound network access from a tool sandbox.

Start a task on a schedule without a new human prompt.

Capture a screen, window, or browser page for use as model input.

Store and provide credentials through a documented secrets feature.

Continue a previously saved product session.

Share a product-defined project workspace with other people.

Show text or tool events as they arrive during a run.

Return JSON or schema-constrained output through a documented product feature.

Define which conversation, instructions, files, memory, and tools a child receives.

Delegate a bounded task to another agent process.

List, inspect, and control child-agent tasks.

Allow a child agent to use MCP tools or declare its own MCP servers.

Control child-agent tools, inherited permissions, and approval behavior.

Return a child agent's findings or work to the parent.

Allow a child agent to create or modify project files.

Propose or execute shell commands in a project environment.

Control or document model-training and service-improvement use of product content.

Inspect documented product usage such as requests, tokens, or cost.

Discover personal or global skills outside the current project.

Upload or select a video for use as model input.

Retrieve live web pages or APIs during a run.

Expose semantic HTML forms as structured tools through WebMCP annotations.

Cancel pending tool registration or execution through abort signals and lifecycle rules.

Register JavaScript-backed tools through document.modelContext.registerTool().

Gate tool exposure across documents and origins through the tools permissions policy.

Communicate read-only and untrusted-content hints across the site-to-agent boundary.

Expose the active page tool list and schemas to a WebMCP-aware agent.

Invoke a page tool with structured arguments and return its result to the agent.

Read or edit files inside a user-selected project.

Search file names, text, or symbols across a workspace.