chess-coach-mcp

A hybrid AI chess coach exposed as an MCP server. A local Stockfish engine supplies grounded evaluations, mistake classifications and tactical motifs; your MCP host model (e.g. Claude) turns those facts into natural-language — Korean or English — coaching.

The engine knows what the best move is. The model explains why. This project wires the two together and adds the piece neither does alone: "here is your recurring mistake, drilled from your own games."

Why this exists

Engines (Stockfish) tell you the best move but not the reason. Raw LLMs explain fluently but play/evaluate chess poorly and hallucinate lines. Existing trainers are fragmented (one app for stats, one for explanations, one for courses) and rarely give a personalised, guided path. chess-coach-mcp is the grounding layer: every claim the coach makes is backed by Stockfish, and the weakness diagnosis is computed from the player's actual games.

Related MCP server: MCP Chess Server

What it does

• Fetch recent games from Lichess or Chess.com (public APIs, no key).

• Analyse a game: per-move classification (best / good / inaccuracy / mistake / blunder, with Korean labels), win% before/after, centipawn loss, the engine's preferred move, tactical-motif tags, and per-phase summary.

• Analyse a position (FEN): top engine lines in SAN, win%, material, and fork/pin flags — so the model can explain the why.

• Diagnose weaknesses across many games: phase weaknesses (opening/middlegame/endgame), recurring tactical blind spots with example positions, leaky openings, a time-trouble proxy, and a ranked top-weakness list.

• Recommend drills: re-solvable positions taken from the player's own blunders, ordered to target their top weaknesses, plus a Lichess daily puzzle warm-up.

What makes it different (차별점)

Most chess tools do one thing well, so you end up stitching several together. chess-coach-mcp is the missing grounding + personalisation layer, delivered right where you already work — inside your AI assistant.

The five things that set it apart:

1. Hybrid & grounded — Stockfish is the judge, the LLM is the explainer. Every coaching claim is backed by the engine, so there are no hallucinated evaluations or made-up lines (the failure mode of asking a raw LLM about chess).

2. Personal, not generic — diagnose_weaknesses aggregates your games into phase weaknesses, recurring tactical blind spots and leaky openings; recommend_drills quizzes you on your own blunder positions — not random puzzles at your rating.

3. Bilingual EN/KO — every structured fact carries a Korean label, so the coaching reads naturally in Korean (한국어 코칭).

4. Local-first & private — your own Stockfish binary + public read-only APIs. No account, no API key, nothing about your games is uploaded anywhere.

5. Inside your assistant — it's an MCP server, so coaching happens in the same chat you already use, composable with everything else your assistant can do.

Tools

source is lichess (default) or chesscom. Every numeric/categorical fact ships with a *_ko Korean label for natural Korean coaching.

Requirements

• Python ≥ 3.11

• Stockfish on your PATH (or set STOCKFISH_PATH):

  • macOS: brew install stockfish

  • Debian/Ubuntu: apt install stockfish

Install & run

uv sync                      # install dependencies
uv run chess-coach-mcp       # run the MCP server over stdio

Register with Claude Code

claude mcp add chess-coach -- uv --directory /ABS/PATH/TO/chess-coach-mcp run chess-coach-mcp

Or add to an MCP client config:

{
  "mcpServers": {
    "chess-coach": {
      "command": "uv",
      "args": ["--directory", "/ABS/PATH/TO/chess-coach-mcp", "run", "chess-coach-mcp"],
      "env": { "STOCKFISH_PATH": "/opt/homebrew/bin/stockfish" }
    }
  }
}

Example coaching flow

"내 리체스 약점 좀 진단해줘. 아이디 myname."

The host calls diagnose_weaknesses("myname"), gets back per-phase ACPL, recurring motifs (e.g. hanging_piece ×4, missed_tactic ×3) with example FENs, then explains in Korean why those positions went wrong and calls recommend_drills("myname") to quiz the user on their own blunders.

Configuration (environment variables)

How move classification works

Moves are classified by the drop in win percentage, not raw centipawns, using Lichess' logistic model — far more meaningful in already-winning or already-losing positions. A move is a blunder if it loses ≥20% win probability, a mistake at ≥10%, an inaccuracy at ≥5%. Mate-aware: a move that throws away a forced mate or walks into one is flagged accordingly.

Tactical motifs are heuristic labels (depth-1 static exchange for hanging pieces, geometric detection for forks/pins/back-rank). They exist to group engine-found mistakes into human themes, not to replace the engine's judgement.

Development

uv run pytest                       # unit + engine tests (skips engine tests if no Stockfish)
uv run python examples/mcp_smoke.py # boot the server over MCP and call tools
uv run python examples/live_check.py <lichess_username>  # live network E2E

Tests marked engine require a Stockfish binary; live tests (none by default) hit the network and are deselected unless you pass -m live.

License

MIT