llm_mcp — Local LLM tools over Model Context Protocol

Exposes [[llm_backends]] as MCP tools so an agent can route work to local models on its own. Pure stdlib, stdio JSON-RPC 2.0, no dependencies.

Tools exposed

Register in Claude Code

Add to .mcp.json at the project root (an example ships at configs/mcp.example.json):

{
  "mcpServers": {
    "botte-llm": {
      "command": "python",
      "args": ["-m", "skills.llm_mcp.server"],
      "cwd": "/absolute/path/to/botte-secrete"
    }
  }
}

Then in a session: "discover my local backends", "classify these 50 tickets locally", "is local routing active?" — the agent calls the tools directly.

Lazy tool loading (default on)

The server has grown to dozens of tools; injecting every full JSON Schema into an agent's context costs thousands of tokens on every turn, whether or not that turn uses them — [[context_profiler]] measured it as the single biggest slice of always-on prefix on this repo. So tools/list returns only a small core (local_chat, auto_route, find_skills, conduct) plus a find_tool(query) meta-tool — the same pattern this very harness uses (ToolSearch). Call find_tool to discover anything else; a strong match returns the full schema in one round trip. tools/call still dispatches any tool by name regardless of what was listed — lazy loading only shrinks the catalog, not what's callable.

Measured saving: ~3.3k tokens (84% of the tool-schema cost). Disable with BOTTE_MCP_LAZY_TOOLS=0 for a client that expects the full catalog upfront.

Manual smoke test

printf '%s\n' \
 '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{}}}' \
 '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
 '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"local_chat","arguments":{"prompt":"ping","max_tokens":16}}}' \
 | python -m skills.llm_mcp.server

Token strategy

Tell the agent: route classification / extraction / short summaries through local_chat, keep architecture & security reasoning on the cloud. That single rule typically moves 30-50% of calls to zero-cost local inference.