The problem with Claude Code PR reviews

I use Claude Code for almost everything. But PR reviews kept frustrating me.

The workflow looks like this without tooling:

1. Find the PR URL
2. Run gh pr diff locally, copy ~800 lines
3. Paste into Claude Code (often truncated mid-function)
4. Ask for a review
5. Get generic feedback: "this looks fine, consider adding error handling"

What's missing? Blame context (who wrote this line, when, why), linked issues (what bug was this PR supposed to fix), and a structured review format that forces citations.

Claude doesn't know any of that by default. You're asking it to review code with no archaeological context.

The fix: a local MCP server

MCP (Model Context Protocol) lets you give Claude Code tools that run locally on your machine. The server speaks JSON-RPC over stdio and Claude calls your tools the same way it calls built-in tools.

I built a Node.js MCP server with 4 tools backed by gh CLI and local git: