metis-ai

Attach a file to the current chat and make it available as a protected download link and preview. Whenever the user asks to attach, send, share, export, or provide a file, you MUST call provide_file; creating the file or mentioning its path is not enough. Call once for each file.

Create a one-time or recurring autonomous agent automation. Every run gets its own persistent chat transcript and can use the same allowed tools, child MCPs, remote tools, and persistent Metis browser as a normal Agent-mode run. Recurring schedules can use minutes, days, or a calendar day each month…

List the current user's automations, including creator, node graph, isolated persistent run chats, status, and recent runs.

Update an existing automation's name, prompt, context chat, models, maximum run duration, timezone, or schedule.

Start an automation immediately as a manual run. The run is isolated in its own persistent chat and does not replace the recurring schedule.

Pause an automation so it will not run until resumed.

Resume a paused automation.

Permanently delete an automation and its run history.

Create or replace a short in-chat checklist. Use only for work with 3+ distinct steps; skip it for a single edit. It appears in the message, not as a side-panel plan. Use create_plan only when the user asked for a plan document.

Create or replace a real Metis plan workspace in the current chat. Provide the complete plan content; a title alone is not sufficient.

Create or replace a real Metis canvas workspace in the current chat. Provide the complete canvas content; a title alone is not sufficient.

Edit an existing plan workspace in the current chat.

Edit an existing canvas workspace in the current chat.

Set the current chat title.

Add concise, non-sensitive keywords to the current chat so it can be found later. Use this silently when the topic becomes clear or changes.

Search the signed-in user's chats by title, keywords, or message content.

Search the owner's personal context hub: devices, servers, services, projects, preferences, and long-term facts. Use before answering questions about the user's setup, before planning changes to their infrastructure, or when background knowledge about the owner would change the answer. Returns groun…

Read the owner's canonical profile: identity, core preferences, infrastructure overview, and active projects. Cheaper than context_search when a general overview is enough.

Store one durable, non-secret fact about the owner in the shared context hub (same fact_id updates the fact). Never store passwords, API keys, cookies, tokens, or auth material. Use sparingly for stable preferences and environment facts the owner confirmed.

List active shared notes in the current chat or workspace scope.

Search shared notes by title or content.

Create a shared note or project for the current chat. Use kind=project with todos for multi-chat planning. Deletion is intentionally not exposed to agents.

Update a shared note using optimistic version checking.

Retrieve the user's saved memories for the current account.

Add a durable memory for the current user. Only save useful user-approved facts or preferences.

Edit one existing memory by id for the current user.

Delete one existing memory by id for the current user.

List plan and canvas workspaces in the current chat.

Delete a plan workspace by id after confirmation.

Delete a canvas workspace by id after confirmation.

Return git status for the agent workspace.

Return the current unstaged git diff for the agent workspace.

Extract readable text from the current browser page and all embedded frames. Use this instead of reloading or shell inspection.

Return structured form controls plus compact text from the page and embedded frames. Controls include a frame hint and stable selector. Prefer this over screenshots or shell/Playwright inspection for forms.

Execute up to 100 browser actions in one round-trip, including embedded-frame controls. Use for repetitive form filling/clicking instead of one tool call per field.

Wait for a selector or visible text condition. Prefer this over fixed wait/sleep calls.

Fill a browser form field using a CSS selector.

Trigger a download from the current browser page by clicking a CSS selector.

Drag an element to a target element (HTML5 and mouse-based drag & drop).

Hover over an element by CSS selector (reveals hover menus and tooltips).

Select an option in a <select> dropdown by value.

Upload a local file into a file input on the current browser page.

List the dynamic MCP registry. Disabled entries show integrations that still need authorization.

Search tools on enabled child MCP servers. Use only when you do not already know the gateway tool name (read_file, edit_file, execute_command, …). Do not call this at the start of ordinary tasks.

List full tool schemas from one child MCP server.

Call any tool exposed by a registered child MCP server.

Required when the current mode cannot perform the user's requested action. Use the exact mode ID: agent for implementation/file changes, plan for read-only planning, or ask for read-only answers; custom mode IDs are supplied in the chat instructions. Opens a real confirmation prompt for switching th…

Pause this agent for a bounded amount of time. Use this when work is expected to finish asynchronously; the wait always has a server-enforced maximum.

Spawn a real provider-neutral Metis child agent with its own persistent chat/run. Use wait=false to launch independent workstreams in parallel, then subagent_status to collect them before final synthesis. The child inherits the parent mode policy, so delegation cannot bypass read-only restrictions.

Inspect delegated subagents for this chat and optionally wait for a status change. Returns compact status data while preserving full tool logs in the chat.

Find an existing matching tool, or autonomously search the complete official MCP Registry, provision and probe the best supported MCP, then return its available tools. Use this when a requested capability is not already visible.

Write or update managed MCP instructions for AGENTS.md, CLAUDE.md, GEMINI.md, Copilot, Cursor, and OpenCode on the server, Windows PC, or laptop.

Add or update a remote URL or stdio MCP server in the dynamic registry. The server remains centrally available through this gateway URL.

Enable or disable one child MCP server.

Search the current web through the local SearXNG instance first, with Exa as fallback. Suitable for current information, products, code, and research.

Fetch and extract one or more public read-only web pages. Uses the local Scrapling static scraper first for speed, then Exa as fallback. If a page requires login, interaction, JavaScript state, or a challenge, use the persistent browser_* tools instead.

Navigate the server-side browser to an allowed URL. The browser can reach server localhost URLs from the configured allowlist.

Return the current server-browser accessibility snapshot for selecting and understanding page elements.

Capture the current server-browser page as a JPEG and return it to the agent. Use only when visual inspection is necessary.

Click a page element by CSS selector, exact/partial visible text, or viewport coordinates. Selectors/text are searched across embedded frames automatically.

Fill a page input by selector or type at the current focus. Selectors are searched across embedded frames automatically.

Press a keyboard key in the server browser.

Set the server-browser viewport resolution for the current tab.

Scroll the current server-browser page vertically.

List and select tabs in the current server-browser session.

Resolve a library name to the best Context7 library id before querying documentation.

Retrieve current, version-specific library documentation from Context7.

Get system overview for the server, CachyOS laptop, or Windows PC.

Run Bash on server/laptop or PowerShell on the Windows PC. Use dedicated Windows UI and Electron tools for visible desktop testing.

Open and control a persistent shell session on a connected remote client. The session keeps its working directory and environment between calls.

List the connected remote clients belonging to the current account, including their IDs, capabilities, policy, and online status.

List a directory on any device.

Read a UTF-8 text file from the server or a connected remote client. Prefer offset+limit around the relevant region instead of reading the whole file. For a client use target client:<remote-client-id>.

Search the persistent local repository index for relevant files, symbols, imports, and paths before broad filesystem exploration.

Return compact indexed codebase findings for a query without dumping whole files.

Prove work claims with real command output instead of asserting them. Each claim pairs a short label with a command and expected output markers; every command is executed and the result recorded in a per-job evidence ledger (see ledger_review). Use for 'tests pass', 'server is up', 'config applied' …

Read the evidence ledger for the current job: every verify_work claim with verified/failed status. Use before reporting task completion so the summary cites real evidence.

Identify or register audio Shazam-style: spectral landmark hashing against a local reference database. Actions: match (identify an audio file), ingest (register a reference track with title/artist), stats (database overview). Requires ffmpeg. Audio argument may be any file ffmpeg can decode.

Find functions, classes, interfaces, types, and exported symbols through the persistent repository index.

Create or overwrite a UTF-8 text file on the server or a connected remote client. For a client use target client:<remote-client-id>.

Replace one exact oldText block with newText in a UTF-8 file on the server or a connected remote client. Pass the smallest unique snippet, never the whole file. For a client use target client:<remote-client-id>.

Delete a file on the server or a connected remote client. For a client use target client:<remote-client-id>.

List Docker containers on a Linux device.

Inspect, restart, start, stop, or read logs for a systemd service on server/laptop.

Inspect and control visible Windows applications through Microsoft's winapp UI Automation. Supports Electron, Win32, WPF, WinForms, and WinUI. Actions include list-windows, inspect, search, status, get-property, get-value, invoke, click, set-value, send-keys, focus, hover, scroll, wait-for, screensh…

Capture a visible Windows app or UI element and return the PNG directly to the agent.

Analyze, install, audit, lint, type-check, unit-test, E2E-test, build, package, launch, or comprehensively validate an Electron project on the Windows PC. Use full first, then launch and windows_ui/windows_screenshot for feature-by-feature visual validation.

Run an arbitrary command inside the active interactive Windows desktop session. Use only when windows_ui or electron_test cannot express the operation.

Save a reusable workflow of allowlisted core gateway tool calls.

List saved workflows.

Get a saved workflow.

Delete a saved workflow.

Run a saved workflow in order, stopping at the first failure. Use dry_run to validate without device side effects.

Return compact live status for all devices, gateway services, and available tool count.

Return the public bootstrap document with connection URL, transport, and discovery tool names.

Search the complete mirrored official MCP Registry for installable servers. Delegates to the registry-autobroker child.

Show official MCP Registry sync state, catalog version, total entries, and knowledge targets.

Return catalog additions, updates, deletions, and provisioning events after a timestamp.