MCP servers

A skillet can attach Model Context Protocol servers to any agent. Each MCP server contributes a set of tools the LLM can call alongside the built-in run_command_line and load_skill_resources. Three transports are supported: stdio (the runtime spawns a subprocess), http, and sse.

Each server is one JSON file. Reference it from .skilled_crew.yaml under the agent that should have access.

Wiring


# .skilled_crew.yaml
agents:
  my_agent:
    instructionsPath: ../my_bot/AGENTS.md
    mcp_servers:
      - filePath: ../dot_claude/mcp_servers/datetime_mcp.json
      - filePath: ../dot_claude/mcp_servers/chrome_devtools.json
    skills: [...]

The path is relative to the .skilled_crew.yaml. Each entry is a single config file. MCP servers are scoped per-agent — only the agent that lists a server gets its tools.

`stdio` config

Subprocess over stdin/stdout. Most MCP servers ship this way.


{
  "name": "datetime_mcp",
  "type": "stdio",
  "command": "npx",
  "args": [
    "tsx",
    "/absolute/path/to/datetime_mcp/src/index.ts"
  ],
  "env": {
    "MY_VAR": "value"
  },
  "cwd": "/optional/working/dir"
}

Field	Required	Purpose
`name`	yes	Display name. Surfaces in `/mcp_servers`.
`type`	yes	`"stdio"`.
`command`	yes	Executable to spawn.
`args`	no	CLI args, defaulting to `[]`.
`env`	no	Extra env vars, merged onto the parent process env.
`cwd`	no	Working directory for the subprocess.

`http` config

Streamable HTTP transport — for a server you run separately and reach over the network.


{
  "name": "my-api",
  "type": "http",
  "url": "http://localhost:3000/mcp"
}

`sse` config

Server-Sent Events transport.


{
  "name": "my-api",
  "type": "sse",
  "url": "http://localhost:3000/sse"
}

Lifecycle

All declared MCP servers are connected before the agent starts handling requests. Stdio servers stay running for the duration of the CLI process. The /mcp_servers built-in slash command in chat lists what’s attached and shows each server’s tools.

Sample servers

Real configs ship under packages/_skillet_agent/data/dot_claude/mcp_servers/:

datetime_mcp.json — local stdio server from the datetime_mcp sibling package.
chrome_devtools.json — npm-published stdio server (chrome-devtools-mcp).

Picking which agent gets which server

Two patterns work in practice:

Everything on the entry agent. Simplest. The orchestrator has every tool, and skills compose with shell commands. Fine until you have many MCP servers and want the LLM to focus.
Servers on the relevant skill’s parent agent in a multi-agent skillet. Heavier setup, but the LLM only sees tools that matter for the current step.

There’s no global “all agents get this server” shortcut — list the server under each agent that should have it.