MCP server

occam-gitignore-mcp exposes the same generator as a Model Context Protocol server. LLM clients can call it as a tool — without inheriting non-determinism from the model itself.

Run

# stdio (default)
uv run occam-gitignore-mcp

# streamable HTTP
uv run occam-gitignore-mcp --transport streamable-http --host 127.0.0.1 --port 8765

# Server-Sent Events
uv run occam-gitignore-mcp --transport sse --port 8765

Tools

occam_gitignore.generate

{
  "tree": ["pyproject.toml", "Dockerfile"],
  "include_comments": true,
  "include_provenance": false,
  "extras": ["secrets/"]
}

Returns the structured GitignoreOutput (content + hash + versions + rules). The LLM never has to invent patterns.

occam_gitignore.diff_against

{ "existing": "*.pyc\n", "tree": ["pyproject.toml"] }

Returns { "added": [...], "removed": [...] }.

occam_gitignore.inspect

{ "tree": ["pyproject.toml", "src/main.py"] }

Returns the fingerprint result (features + evidence) — useful when an LLM wants to reason about which features were detected before committing to a generation.

Why MCP for a deterministic tool?

The MCP adapter exists precisely because the rest of the stack is deterministic. It gives an LLM a way to delegate the part of the task that should never be guessed — the actual ignore patterns — to a verifiable function, while keeping the LLM in charge of the parts that legitimately need judgement (which user extras to add, which paths to scan, how to explain the result to the user).