Use open-nitpick from an agent

nitpick mcp serves local review tools over MCP standard input and output. It uses the configuration and provider environment of the repository where it runs. It does not publish pull-request comments; the agent receives the result and decides what to do with it.

Install a client entry

See supported clients and where each client stores a project or user entry:

nitpick mcp clients

Preview a project-scoped entry before writing it:

nitpick mcp install claude-code -print

Then write it:

nitpick mcp install claude-code

To register the server with every client found on this machine at once, run nitpick setup (nitpick setup -dry-run lists the files first).

-print writes nothing. Project entries vary by client: Claude Code uses .mcp.json, Cursor uses .cursor/mcp.json, VS Code uses .vscode/mcp.json, OpenCode uses opencode.json, and Gemini CLI uses .gemini/settings.json. Claude Desktop, Windsurf, and Codex have user entries only. Use -user where a client supports a user-scoped entry. nitpick mcp install <client> -h is the current source of truth for each location and limitation.

Reinstalling updates Nitpick's command and arguments while preserving custom server settings, including environment variables, timeouts, and disabled state. Other servers and settings are kept.

For Codex, installation adds tool_timeout_sec = 3600 when the Nitpick entry has no explicit timeout. Reinstalling keeps your configured value and TOML comments outside the values it updates. Restart the client to load the new settings. The client timeout limits the whole tool call; Nitpick's model timeout limits each model request, and a review can make several requests. Increase the client timeout for longer reviews, or run nitpick review -dry-run in a tracked local process to avoid the MCP wait limit.

A review call has a total server deadline of 30 minutes. Set timeout_seconds to 1–86400 to change it (0 uses the default), and keep the client timeout longer than that deadline. Clients that supply an MCP progress token receive stage messages and a heartbeat every 30 seconds, including elapsed time. A heartbeat says the request is still pending; it does not prove model progress. Clients without a progress token still receive the result or error. Cancellation and the total deadline propagate to model requests and credential commands. A deadline error is a failed review, not a clean result.

Tools and returned data

The server exposes review, full_review, repo_score, code_smell, ai_slop, security_scan, and explain_config. review returns a summary, findings with path, line, severity, class, title and rationale, analyzer outcomes, the gate result, and complete/failed_stages. Whole-tree tools add covered, unreviewed, unbudgeted, and skipped files; repo_score also adds a scorecard.

Use review for a changed working tree. Use full_review, repo_score, or code_smell with paths to bound a broader inspection. security_scan reports whether required stages completed, and ai_slop supports no_model for its deterministic checks. Start the transport directly only when your client needs a manual command:

nitpick mcp -repo .

The process reserves standard output for MCP. Use -v for logs on standard error. All workflows explains how the same configuration works outside an agent session.