Setup Doctor MCP server

setup-doctor

Score and improve your AI coding agent setup. Local-only, open source, free.

That's this repository's own real score, produced by running setup-doctor against itself (see docs/notes.md for how). It is not a mockup.

setup-doctor audits how Claude Code, GitHub Copilot CLI, Codex or Cursor is configured in your project and globally, gives it a 0-100 score with concrete fixes, and turns your local usage logs into a shareable "Wrapped" card. Everything runs on your machine. Nothing is ever sent anywhere.

This is real npx setup-doctor output (against a small demo project, not this repository): a critical secret finding, a dead MCP server command, a thin skill description, and the score capped because of the critical finding. Not a mockup.

What it does

• Doctor: audits instruction files (CLAUDE.md / .github/copilot-instructions.md / AGENTS.md / .cursorrules), skills, subagents, MCP servers, plugins, settings and hooks. Runs 29 rules across 6 categories and returns a score, a band (Excellent / Good / Needs work / Poor), and a specific fix for every finding.
• Wrapped: summarizes your local session logs (Claude Code, Codex, GitHub Copilot CLI or Cursor) (sessions, active days, tokens, an API-equivalent cost estimate, streaks, busiest hour, a persona label) into a shareable card, in three visual themes and two sizes, with optional PNG export.
• Badge: a static or live (shields.io endpoint) README badge showing your current score.
• HTML report: a single self-contained, themed report file. No network requests, strict Content-Security-Policy, everything inlined.
• MCP server mode: npx setup-doctor mcp exposes Doctor and Wrapped as read-only tools over stdio for Claude Desktop and other MCP clients; see docs/guide/mcp-server.md. Listed on the official MCP Registry as io.github.saxenakapil/setup-doctor.

Supported agents

Subagents, plugins, settings and hooks checks are Claude Code/Copilot-specific; those categories are simply excluded from the score for Codex/Cursor-only setups rather than counted against you. Skills checks (.claude/skills for Claude Code/Copilot CLI, project-scope .cursor/skills for Cursor) apply to all three; Cursor's own global/personal skills live in a cloud-synced store rather than a fixed local path, so only its project-scope skills are read (see docs/notes.md). Copilot CLI is documented to read several of Claude Code's own files directly (.claude/skills, .claude/settings.json, and the "portable format" .mcp.json); when a project is detected as both agents, a real problem in one of those shared files is scored once, not once per agent, and the report notes which other agent it also affects.

Install and run

No install needed, run it with npx:

Shortened. The full README is on GitHub.