MCP Servers

MCP servers

This agent has a built-in MCP client. Servers are declared in mcp.json; their tools appear as mcp_<server>_<tool> once connected. The model can also always find and call MCP tools via a single mcp proxy tool (mcp({ search: "..." }) / mcp({ tool: "...", args }) without their schemas ever being loaded into context, and without a lifecycle: "lazy" server being connected until one of its tools is actually needed — see "Context window control" below.

Config files

Project entries override global entries with the same server name; project settings keys override global ones per key. Prefer project-local config for project-specific servers, global for personal/general-purpose ones. Create the file if it doesn't exist.

Applying changes: config is read at session start. After editing mcp.json, tell the user to run /reload (this re-reads config, restarts servers, and refreshes tools). You cannot run /reload yourself.

Config format

{
  "settings": {
    "toolPrefix": "mcp",
    "requestTimeoutMs": 30000,
    "maxRetries": 3,
    "searchResultLimit": 15
  },
  "mcpServers": {
    "<server-name>": { ... }
  }
}

settings is optional. Server names: keep them short and lowercase; they become part of tool names.

Local (stdio) server — spawned as a subprocess

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"],
      "env": { "SOME_VAR": "literal-value" }
    }
  }
}

• command is required; transport defaults to "stdio".
• env values are literals merged over the parent environment — there is no ${VAR} interpolation. If a server needs a secret, ask the user to provide it or reference their shell environment by launching via a wrapper script.
• Most published servers run via npx -y <package> (Node) or uvx <package> (Python). If unsure of the package name, search the web for "<product> MCP server".

Remote (HTTP) server

{
  "mcpServers": {
    "internal": {
      "transport": "streamable-http",
      "url": "https://mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer <api-key>" }
    }
  }
}

• transport: "streamable-http" (preferred) or "sse" (legacy); url required.
• headers is for static API-key auth. Never invent keys — ask the user for theirs.

Remote server with OAuth (user login in browser)

{
  "mcpServers": {
    "linear": {
      "transport": "streamable-http",
      "url": "https://mcp.linear.app/mcp",
      "auth": { "type": "oauth" }
    }
  }
}

• auth: { "type": "oauth" } is enough for most servers (discovery + dynamic client registration + PKCE are automatic). Optional fields: scope, clientId, clientSecret, redirectUrl (only for pre-registered clients), clientName.
• After /reload, the first connection will fail with "Authentication required" — that is expected. Start the login yourself with the mcp_auth tool (opens the user's browser), or tell the user to run /mcp:auth <server-name>. Tokens are stored and refreshed automatically afterwards.
• OAuth is only valid for streamable-http/sse, not stdio.
• If dynamic client registration is rejected (e.g. a client-name policy error), set "clientName" in the auth object and retry.

Other per-server options

• lifecycle: "lazy" (default — starts on first use of one of its tools via the mcp proxy tool, or manually with /mcp:start <name>) or "eager" (starts at session start). Use eager only for a server you want connected from the very first turn.
• requestTimeoutMs, healthCheckIntervalMs: numeric overrides, rarely needed.
• idleTimeoutMs: lifecycle: "lazy" only — auto-disconnect this many ms after the server's last tool call. Good for servers used in bursts, e.g. 600000 (10 min).
• description: one-line summary shown by mcp search before this server has ever connected (its real tool list isn't known yet). Set this on lazy servers so the model can find them via search before the first connection.
• directTools: false (default — all tools stay searchable-only via mcp), true (all tools load straight into context), or an array of MCP-side tool names to keep direct while the rest stay proxy-only. Set true (or list specific names) only for a small server used on nearly every turn, where a search round-trip isn't worth it.