The GitHub MCP server is GitHub’s own server, maintained in the github/github-mcp-server repository. It gives Claude read and (optionally) write access to your GitHub repositories: code search, issue management, PR review, and more. GitHub hosts a remote version at https://api.githubcopilot.com/mcp/, and you can also run it locally with Docker.
The older Anthropic reference server for GitHub is archived and its npm package is deprecated, so use GitHub’s server instead. Because GitHub is central to dev workflows, it is a natural first install for engineers and technical founders.
Why use it
GitHub is where most engineering work lives. Adding it as an MCP server collapses dozens of tab-switches per day into single prompts. “Show me the PRs that touched our auth middleware in the last week” goes from a manual GitHub UI tour to a one-line answer.
For solo founders running Claude Code, the GitHub server is what makes “fix this issue” workflows feel native. Claude reads the issue, searches the relevant code, opens a PR, and links them together without leaving the conversation.
What it actually does
A wide surface, grouped into toolsets: repositories and code, issues, pull requests, GitHub Actions workflows, code security and more. You can turn toolsets on or off to keep the tool list small.
Practical patterns:
- “Find all open PRs in our org that have been waiting for review for more than three days.”
- “Open an issue in repo X titled ‘fix sitemap lastmod bug’, label it ‘bug’ and assign to @david.”
- “Search every repo I own for usages of the deprecated useEffectAsync hook.”
Gotchas
Scope your access narrowly. A token with write permissions can do a lot of damage if Claude misinterprets a prompt. Default to read-only and grant write only when you need it. Fine-grained personal access tokens let you limit access to chosen repositories.
Rate limits apply. GitHub’s API caps at 5,000 requests per hour per token. Heavy queries (e.g. searching across hundreds of repos) can chew through this quickly. If you hit the limit, the server will return clear errors but Claude may not always interpret them gracefully.
Combine with Sentry for an end-to-end debug flow: error in production → Sentry pulls the stack trace → GitHub finds the introducing PR → Claude drafts a fix.
