decisionlint-mcp

An MCP server that lints the questions a project sends to TypeSafe's Jev model or to OpenAI's Decisions API (gpt-6-luna) before they are sent, and mends the ones it can. It is the engine behind the DecisionLint editor extension, as eight tools an agent can call.

Linting needs no API key and makes no network calls, and nothing here writes a file. Two tools ask the model itself, with your key, when you ask them to.

Use it

Point any MCP host at npx decisionlint-mcp.

Claude Code

claude mcp add decisionlint -- npx -y decisionlint-mcp

Anything with a JSON config, such as Cursor, Windsurf or Claude Desktop:

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

VS Code needs nothing here. Install the extension instead. It carries this server and registers it for you: VS Code Marketplace · Open VSX

Already running decisionlint in CI? The same server is npx decisionlint --mcp. This package is the one with no flag to forget: a client that starts decisionlint bare gets a linter that exits, not a server.

Prefer a global install to npx on every launch:

npm install -g decisionlint-mcp

{
  "mcpServers": {
    "decisionlint": { "command": "decisionlint-mcp" }
  }
}

No configuration of its own, and no key unless you want a check with the model. To check it before wiring it into anything:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"check","version":"0"}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | npx -y decisionlint-mcp

Tools

Tool	What it does	Sends
lint_text	Lints a request body or source code passed as text. Each finding carries the edit that would mend it	nothing
lint_paths	Lints files or directories on disk	nothing
fix_text	Returns the text with the safe fixes applied, and the report of what is left. Writes nothing	nothing
list_rules	Lists every rule with what it means, its default level and its pages	nothing
explain_rule	A rule's own page as Markdown: an example that is flagged, one that is not, how to fix it and how to silence it	nothing
plan_ask	What ask_model would send for the text or files: requests, model, tokens, questions held back	nothing
ask_model	Asks a decision model to check the questions in text or files, and returns the lint report with the model's findings added. platform names where: typesafe for Jev, openai for Luna, openrouter, vercel, llmgateway or liquid, each with its own key variable	the questions, with your key
probe_question	Sends one question as written and with its layout varied, and reports whether the answer held	about a dozen requests, with your key

Every tool answers as structured content beside its text. The first six are read-only and offline, so a host can let an agent call them without asking. The two that send are marked open-world and not idempotent, each call costs, and their descriptions tell an agent to call them only when you ask for a check or a probe. lint_paths and ask_model read the files under the paths they are given and nothing else. A decisionlint.json near those files applies as it does on the command line, and rules on any call overrides it.

The key

ask_model and probe_question look for TYPESAFE_API_KEY, or OPENAI_API_KEY for Luna, in this order, and use the first they find:

1. The environment the client started the server with. Most clients take an env map beside command in their config.
2. .env.local, then .env, in the directory the client started the server in, which is the project for every client that runs servers per project.

A key is never taken as a tool argument, so none can land in an agent's transcript or a host's log, and no answer ever carries one: the report says where the key came from, not what it was. With no key anywhere the call fails before anything is sent and says where it looked.

It is listed on the MCP registry as io.github.nolindnaidoo/decisionlint.

License

MIT