llm_mcplisted

# 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 | Tool | Purpose | |------|---------| | `discover_backends` | Scan localhost/network, register reachable LLM servers | | `list_models` | List models the registered backends expose | | `audit_local_usage` | Local-usage report + hardware-aware setup steps | | `route_task` | Recommend cheapest tier + a concrete local backend/model | | `local_chat` | Run a prompt on a local model (0 cloud tokens) | | `auto_route` | Decide and optionally execute the cheapest capable route | | `route_feedback` | Verify an executed route by its `feedback_id` | | `qa_status` | Show verified quality-memory maturity and its next step | | `qa_advise` | Explain a shadow-only k-NN route suggestion or abstention | | `qa_record` | Add an externally verified outcome without raw task storage | | `compress` | Compress context; reversible calls add deduplicated verified grounding labels | ## Register in Claude Code Add to `.mcp.json` at the project root (an example ships at `configs/mcp.example.json`): ```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 ac