stockfish-mcp

A tiny, opinionated Model Context Protocol server that gives an LLM the best move for any chess position — powered by Stockfish.

Hand it a position as FEN or PGN (or nothing, for the starting position) and it returns the best move, the engine's evaluation, and — optionally — the predicted best line, as clean structured JSON. One tool. One call. No UCI knowledge required.

{
  "turn": "black",
  "bestmove": { "uci": "f6e4", "san": "Nxe4" },
  "ponder":   { "uci": "d2d4", "san": "d4" },
  "score": { "type": "cp", "value": -371, "perspective": "side to move" },
  "evaluation": "White is better (+3.71)",
  "fen": "r1bqk2r/ppp2ppp/2np1n2/P3p3/1PB1P3/5N2/2PP1PPP/RNBQ1RK1 b kq - 0 7"
}

Why another Stockfish MCP?

Most chess-engine MCP servers are thin UCI passthroughs: they hand the language model the raw UCI protocol (position, go, setoption, session IDs) and expect it to orchestrate a stateful, multi-call conversation with the engine. That's the wrong abstraction for an LLM, and it's a source of constant confusion — empty "silent command" responses, position state lost between calls, malformed FENs, and raw dumps of 30 search-trace lines.

stockfish-mcp takes the opposite approach. The model gets one high-level operation and the server handles everything else:

• One tool, one call. Set the position and search it in a single analyze call. No position-then-go dance, no sessions to manage.

• FEN or PGN in. Paste a FEN, a full PGN, a bare move list, or startpos. Short FENs missing the halfmove/fullmove fields are accepted and normalized.

• Unambiguous evaluation. UCI reports scores from the side-to-move's perspective (a classic footgun). We normalize to a White-perspective summary — "White is better (+3.71)", "Mate in 1 for Black" — that a model can relay verbatim without misreading who's winning.

• Human-readable moves. Best move and ponder move in SAN (Nxe4, O-O, e8=Q+), alongside the raw UCI.

• Optional best line. includeLine: true adds the engine's principal variation as a numbered SAN line (1. e4 c5 2. Nf3 …), truncated for readability. Off by default to keep responses tight.

• Stateless & robust. A fresh engine process per call (startup is ~tens of ms vs. a multi-second search) means no shared state, no race conditions, no leaked startup banners.

• Clear errors. Terminal positions (checkmate/stalemate/draw) and invalid input return descriptive messages instead of hanging or crashing.

Requirements

• Node.js ≥ 18

• Stockfish installed and on your PATH (or point STOCKFISH_PATH at the binary). This project does not bundle Stockfish.

Install Stockfish:

# Debian / Ubuntu
sudo apt install stockfish

# macOS (Homebrew)
brew install stockfish

# or download a binary from https://stockfishchess.org/download/

Usage

With any MCP client (Claude Desktop, etc.)

Add to your client's MCP config. If published to npm:

{
  "mcpServers": {
    "stockfish": {
      "command": "npx",
      "args": ["-y", "stockfish-mcp"]
    }
  }
}

Or run it from a local checkout:

{
  "mcpServers": {
    "stockfish": {
      "command": "node",
      "args": ["/path/to/stockfish-mcp/index.mjs"]
    }
  }
}

From the command line (manual testing)

npm install
npm start          # speaks MCP over stdio
npm test           # end-to-end smoke test (needs stockfish installed)

The analyze tool

Response

Examples

FEN (black to move):

// analyze({ "position": "r1bqk2r/ppp2ppp/2np1n2/P3p3/2B1P3/5N2/2PP1PPP/RNBQ1RK1 b kq - 0 7", "depth": 14 })
{
  "turn": "black",
  "bestmove": { "uci": "f6e4", "san": "Nxe4" },
  "ponder": { "uci": "d2d4", "san": "d4" },
  "evaluation": "White is better (+3.71)"
}

PGN:

// analyze({ "position": "1. e4 e5 2. Nf3 Nc6 3. Bb5 a6", "depth": 12 })
{
  "turn": "white",
  "bestmove": { "uci": "b5c6", "san": "Bxc6" },
  "ponder": { "uci": "d7c6", "san": "dxc6" },
  "evaluation": "White is better (+0.36)"
}

Best line + mate detection:

// analyze({ "position": "6k1/5ppp/8/8/8/8/8/R6K w - - 0 1", "includeLine": true })
{
  "turn": "white",
  "bestmove": { "uci": "a1a8", "san": "Ra8#" },
  "evaluation": "Mate in 1 for White",
  "line": "1. Ra8#"
}

Configuration

Environment variables:

Example:

{
  "mcpServers": {
    "stockfish": {
      "command": "npx",
      "args": ["-y", "stockfish-mcp"],
      "env": { "STOCKFISH_PATH": "/usr/games/stockfish", "STOCKFISH_DEFAULT_DEPTH": "20" }
    }
  }
}

How it works

The server spawns a fresh stockfish process for each analyze call, sends position fen … followed by go depth N (or go movetime N), reads the engine's info lines to capture the latest score and principal variation, and resolves on bestmove. Moves are converted to SAN with chess.js, which also handles PGN→FEN conversion and FEN validation.

Two UCI subtleties worth noting (both handled here): go is asynchronous, so stdin is left open and no quit is sent before the search finishes — closing the pipe would abort it. And score cp/score mate are reported from the side-to-move's perspective, so we flip them when Black is to move to produce a consistent White-perspective evaluation.

License

MIT.

A note on Stockfish: Stockfish itself is licensed GPL-3.0. This project does not include, link against, modify, or distribute Stockfish — it only spawns a Stockfish binary that you install separately and communicates with it over the public UCI protocol via a pipe. The two are separate programs, so this wrapper is independently licensed under MIT. You are responsible for obtaining Stockfish and complying with its license.