MCP Tools — The Brain

Overview

The Brain exposes itself as a Model Context Protocol server via mcp_server.py. This allows Claude to call the delegate tool natively — it appears in Claude's tool list alongside Bash, Read, and Edit — rather than via a Bash call. The same orchestrator runs in both modes.

Registration

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "the-brain": {
      "command": "python",
      "args": ["C:\\Claude\\git\\the-brain\\mcp_server.py"]
    }
  }
}

Claude Code picks it up automatically on next launch. Run python mcp_server.py from the command line to verify the server starts before registering.

Exposed Tool: `delegate`

This is the only tool The Brain exposes. Claude reaches for it automatically for any AI task that matches the routing guide in CLAUDE.md.

Input Schema

Parameter	Type	Required	Description
`provider`	string	yes	Provider key: `cerebras` `groq` `gemini` `mistral` `sambanova` `fireworks` `openai` `huggingface` `cloudflare` `openrouter` `ollama` `pollinations`
`task_type`	enum	yes	`classification` `summarization` `coding` `creative` `reasoning` `factual_qa` `extraction` `translation` `general`
`prompt`	string	yes	Full prompt to send to the worker
`context`	string	no	Optional context block prepended to the prompt
`max_tokens`	integer	no	Maximum output tokens (default: 2048)

Response Format

Each call returns TextContent with a header line followed by the content:

[mistral / mistral-small-latest | 512 tokens | 1172ms | free]

<content from the worker>

CLI Equivalent

The MCP tool call above is equivalent to running:

python delegate.py --provider mistral --type coding --prompt "..."