CodeRabbit MCP

MCP (Model Context Protocol) server that exposes CodeRabbit IDE extension review comments as tools and resources. Use it from Cursor or any MCP client to scan the latest batch, list unaddressed comments, read exported markdown by file/kind, and mark comments as addressed or wontfix.

The comments this server reads are written locally by the CodeRabbit IDE extension (VS Code / Cursor). Install that extension and run a CodeRabbit review in your workspace; this MCP server then discovers the latest batch from the extension’s storage and exposes it as tools and resources.

• Node: ≥20
• Transport: stdio (stdin/stdout for JSON-RPC; stderr for logs)

Installation

This server is stdio only (no remote URL). Use the local connection config for your client.

Go to: Settings → Cursor Settings → MCP → Add new global MCP server

Pasting the following configuration into your Cursor ~/.cursor/mcp.json file is the recommended approach. You may also install in a specific project by creating .cursor/mcp.json in your project folder. See Cursor MCP docs for more info.

Since Cursor 1.0, you can click an install button below for instant one-click installation.

Local Server Connection

{
  "mcpServers": {
    "CodeRabbit MCP": {
      "command": "npx",
      "args": ["-y", "@deffz-finesse/coderabbit-mcp"]
    }
  }
}

Optional: add "env": { "RABBIT_DISCOVERY_ROOT": "/path/to/workspaceStorage", "RABBIT_OUTPUT_BASE": "/path/to/output" } to override discovery or output paths.

Run this command. See Claude Code MCP docs for more info.

Local Server Connection

claude mcp add CodeRabbit-MCP -- npx -y @deffz-finesse/coderabbit-mcp

Open Claude Desktop developer settings and edit your claude_desktop_config.json file. See Claude Desktop MCP docs for more info.

Local Server Connection

• macOS / Linux: ~/.claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "CodeRabbit MCP": {
      "command": "npx",
      "args": ["-y", "@deffz-finesse/coderabbit-mcp"]
    }
  }
}

Restart Claude Desktop after editing.

Add this to your VS Code MCP config file, or click the Install in VS Code badge at the top for one-click install (opens VS Code). See VS Code MCP docs for more info.

Local Server Connection

"mcp": {
  "servers": {
    "CodeRabbit MCP": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@deffz-finesse/coderabbit-mcp"]
    }
  }
}

Add this to your Antigravity MCP config file. See Antigravity MCP docs for more info.

Local Server Connection

{
  "mcpServers": {
    "CodeRabbit MCP": {
      "command": "npx",
      "args": ["-y", "@deffz-finesse/coderabbit-mcp"]
    }
  }
}

Add the following to your client’s MCP config. Use mcpServers (or your client’s equivalent) with the same command and args.

{
  "mcpServers": {
    "CodeRabbit MCP": {
      "command": "npx",
      "args": ["-y", "@deffz-finesse/coderabbit-mcp"]
    }
  }
}

Install (from source)

npm install -g @deffz-finesse/coderabbit-mcp

Or clone and build:

git clone <repo-url>
cd coderabbit-mcp
npm install
npm run build

Run (manual)

The server uses stdio (stdin/stdout). When not using the Installation snippets above: command = node (or npx), args = path to dist/index.js or ["-y", "@deffz-finesse/coderabbit-mcp"]. It does not listen on a port.

Environment

|Variable|Purpose| |------|--------| |RABBIT_DISCOVERY_ROOT|Root directory to search for CodeRabbit batch folders. Default: ~/.cursor-server/data/User/workspaceStorage (Cursor workspace storage). Set this for VS Code or other editors to point at their workspace storage.| |RABBIT_OUTPUT_BASE|Base directory for the SQLite DB and exported markdown. Default: .coderabbit under the current working directory, then ~/.coderabbit if cwd is not writable.|

Data written by the server:

• DB: <RABBIT_OUTPUT_BASE>/rabbit-comment.db (or .coderabbit/rabbit-comment.db when using default).
• Exports: <RABBIT_OUTPUT_BASE>/exports/rabbit-<batchId>/ (markdown files per file/kind).

Available tools

All tools are under the rabbit namespace.

|Tool|Description| |------|-------------| |rabbit.ping|No-op; use to check the server is reachable.| |rabbit.scan|Discover latest CodeRabbit batch, parse comments, export markdown, persist to DB. Returns JSON with ok, batchId, exportDir, filesProcessed, commentsExtracted, unaddressedCount (or error on failure). No parameters; idempotent.| |rabbit.list_unaddressed|List unaddressed comments for a batch. Optional batchId (default: latest). Returns byFile, countByKind, resourceUris per file.| |rabbit.get_comment|Get full comment record and status history by comment_id (fingerprint). Requires comment_id.| |rabbit.mark_addressed|Set comment status to addressed. Requires comment_id; optional note.| |rabbit.mark_wontfix|Set comment status to wontfix. Requires comment_id; optional note.|

Resources

• rabbit-markdown: Exported markdown for a repository file and comment kind within a batch. URI form: rabbit://b/{batchId}/f/{filename}/{kind} (filename and kind are URL-encoded). List returns URIs for the latest batch; read returns the markdown content.

Prompts

• rabbit-triage: Runs a scan, lists unaddressed comments by file, and returns a triage summary with suggested next steps.

For detailed installation steps (Cursor, VS Code, Claude, etc.), architecture, configuration, and development, see the project docs.

Develop

npm install
npm run dev          # Run with tsx (stdio; connect an MCP client)
npm run build        # Compile TypeScript to dist/
npm run typecheck    # tsc --noEmit
npm run lint:check   # Biome check
npm run lint:fix     # Biome check --write
npm test             # Build + Node test runner

Security / Trust model

This server is local-only: it runs as a subprocess of your MCP client (e.g. Cursor) and does not open network ports or authenticate users. All data stays on your machine.

• No authentication: Anyone who can run the server can run all tools and read all resources. Use normal OS and editor access controls.
• Paths: Discovery root (RABBIT_DISCOVERY_ROOT) and output base (RABBIT_OUTPUT_BASE) are taken from the environment. The server reads from the discovery root and writes the DB and exports under the output base. Ensure these paths point at locations you trust.
• Inputs: Tool and resource inputs (e.g. comment_id, batchId) are used only as identifiers; path traversal is prevented when reading exported markdown. Tool arguments are validated with Zod; invalid args return a structured error (Invalid arguments with details) rather than being passed through.

License

MIT © Deffz Finesse. See LICENSE for the full text.