Tool Reference

mcp_discover

Search tool catalogs across all child MCP servers. Without a query, returns server status and tool counts.

Parameters

Parameter	Type	Required	Description
`query`	string	no	Search query for tools
`server`	string	no	Filter to a specific server

Example 1: List all servers

Request (no parameters):

{}

Response:

[
  {
    "name": "playwright",
    "state": "idle",
    "toolCount": 52,
    "criticality": "vital"
  },
  {
    "name": "sqlite",
    "state": "idle",
    "toolCount": 6,
    "criticality": "vital"
  }
]

Example 2: Search for a capability

{ "query": "screenshot" }

Response:

[
  {
    "tool": "browser_screenshot",
    "server": "playwright",
    "description": "Take a screenshot of the current page",
    "score": 12,
    "confidence": 0.95
  }
]

Example 3: Filter to a specific server

{ "server": "playwright" }

Returns the status of only the playwright server and its tools.

mcp_provision

Intent-based provisioning. Describe what you need, and MetaMCP resolves and optionally provisions the right server.

Parameters

Parameter	Type	Required	Description
`intent`	string	yes	What capability you need
`context`	string	no	Additional context for resolution
`autoProvision`	boolean	no	Auto-provision if trusted (default: false)

Example 1: Find a server for a capability

{ "intent": "I need to crawl a website and extract links" }

Response (local match):

{
  "source": "local",
  "tools": [
    {
      "tool": "browser_navigate",
      "server": "playwright",
      "description": "Navigate to a URL",
      "confidence": 0.87
    }
  ]
}

Example 2: With context

{
  "intent": "database access",
  "context": "I need to query a PostgreSQL database with read-only access"
}

The context string is appended to the intent for more precise matching.

Example 3: With autoProvision enabled

{
  "intent": "I need to interact with GitHub issues",
  "autoProvision": true
}

Response (registry match, trusted):

{
  "source": "registry",
  "matches": [
    {
      "name": "@modelcontextprotocol/server-github",
      "description": "GitHub API integration",
      "confidence": 0.92,
      "trusted": true,
      "autoProvisionable": true,
      "installCommand": "npx -y @modelcontextprotocol/server-github"
    }
  ]
}

If autoProvision is false (the default), the trusted and autoProvisionable fields are omitted.

mcp_call

Forward a tool call to a specific child MCP server. Retries once on crash for vital servers.

Parameters

Parameter	Type	Required	Description
`server`	string	yes	Target server name
`tool`	string	yes	Tool name to call
`args`	object	no	Arguments to pass to the tool

Example 1: Simple call

{
  "server": "playwright",
  "tool": "browser_navigate",
  "args": { "url": "https://example.com" }
}

Example 2: Call with complex arguments

{
  "server": "sqlite",
  "tool": "query",
  "args": {
    "sql": "SELECT name, email FROM users WHERE created_at > ? ORDER BY created_at DESC LIMIT 10",
    "params": ["2024-01-01"]
  }
}

Example 3: Call with no arguments

{
  "server": "playwright",
  "tool": "browser_screenshot"
}

When args is omitted, the tool is called with no arguments.

mcp_execute

Execute JavaScript code in a V8 sandbox with access to all provisioned servers. Supports async/await, sleep(ms), and console.log.

Parameters

Parameter	Type	Required	Description
`code`	string	yes	JavaScript code to execute

Sandbox limits

Limit	Value
Max code size	50KB
Execution timeout	120,000ms (120s)
Max sleep per call	30,000ms (30s)
Max output size	10MB

Example 1: Simple single-server call

{
  "code": "const result = await servers.sqlite.call('query', { sql: 'SELECT count(*) FROM users' }); return result;"
}

Example 2: Multi-server composition

{
  "code": "await servers.playwright.call('browser_navigate', { url: 'https://example.com' }); const screenshot = await servers.playwright.call('browser_screenshot', {}); const dbStatus = await servers.sqlite.call('query', { sql: 'SELECT 1' }); return { screenshot, dbStatus };"
}

Example 3: With sleep and console output

{
  "code": "console.log('Navigating...'); await servers.playwright.call('browser_navigate', { url: 'https://example.com' }); await sleep(2000); console.log('Taking screenshot...'); const result = await servers.playwright.call('browser_screenshot', {}); return result;"
}

Response includes captured console output prepended to the return value:

Navigating...
Taking screenshot...
{ ... screenshot data ... }

The sandbox provides: JSON, Math, Date, Array, Map, Set, Promise, Object, String, Number, Boolean, RegExp, Error, parseInt, parseFloat, isNaN, isFinite, encodeURIComponent, decodeURIComponent, encodeURI, decodeURI, setTimeout, clearTimeout, console, sleep, and the servers Proxy.

Not available: process, require, module, exports, Buffer, __dirname, __filename, eval (throws EvalError), Function constructor, SharedArrayBuffer, WebAssembly, fetch, globalThis.

Response format

All tools return MCP-standard content arrays with type: "text" entries. The text content is typically JSON-formatted.

{
  "content": [
    {
      "type": "text",
      "text": "{ ... JSON response ... }"
    }
  ]
}

Error handling

When a tool call fails, the response includes isError: true and an error message in the content.

{
  "content": [
    {
      "type": "text",
      "text": "Error calling browser_navigate on playwright: Connection refused"
    }
  ],
  "isError": true
}

The LLM can inspect the error and retry, adjust arguments, or try a different approach.

For mcp_execute, sandbox errors include the prefix Sandbox error: or mcp_execute error: in the message.

PreviousClaude Code Setup NextCLI Reference