GitHub and AI: Is there an MCP server for GitHub?

GitHub hosts source code and guides teams through issues, pull requests and builds. If you want an AI assistant such as Claude or Copilot to work with this data, the first question is: is there an MCP server for it?

In short

Yes. GitHub maintains an official MCP server, hosted and local. According to the changelog, the hosted server has been generally available since 4 September 2025. According to the documentation, it is open to all GitHub users, regardless of plan. Individual tools need the same license as the matching GitHub feature.

What GitHub offers itself

The hosted server. According to the repository, it runs at https://api.githubcopilot.com/mcp/. According to the repository, GitHub builds it with the repository as a library and updates it regularly. It has additional tools, for example create_pull_request_with_copilot, which starts the Copilot coding agent.

The local server. The repository is MIT-licensed. The server runs as the Docker image ghcr.io/github/github-mcp-server or as a binary.

Sign-in. According to GitHub Docs, the hosted server supports OAuth and personal access tokens (PAT). With OAuth, the server sees only the scopes you approve at sign-in. OAuth requires the host to have registered a GitHub App or OAuth App. According to the support table, this is not the case for Claude Code and Claude Desktop: they use the hosted server with a PAT only. The Claude Code documentation shows the PAT as the way. The guide for Claude says the remote server is not supported as a custom connector in Claude Desktop and recommends the local server. The Claude Desktop documentation for third-party inference also names a PAT for the remote connector. According to the repository, the local server signs in with OAuth in the browser or with a PAT.

Toolsets and read-only mode. Toolsets group the tools, for example repos, issues, pull_requests or actions. Without a setting, the README applies a default set of context, repos, issues, pull_requests and users. Read-only mode (X-MCP-Readonly, the URL suffix /readonly or --read-only) disables every tool that does not only read.

A REST API. The GitHub REST API works, according to the documentation, with personal access tokens, tokens from a GitHub App or the GITHUB_TOKEN in GitHub Actions.

AI in the product. GitHub Copilot suggests code, answers questions and works through tasks. For Copilot Free, Pro and Pro+, GitHub uses interaction data for training from 24 April 2026 unless users opt out, according to the blog. According to GitHub, Copilot Business and Copilot Enterprise are not affected. We found no statement in the blog post on Copilot Max and Copilot Student, which the plans page also lists. According to the Generative AI Services Terms (for direct purchase), GitHub does not use inputs or outputs for training unless customers give documented instructions. According to the plans page, Copilot is not available on GitHub Enterprise Server.

Data location. According to GitHub, GitHub stores data for GitHub.com in the USA by default. With GitHub Enterprise Cloud with data residency, you choose the location: EU (with Azure regions in Norway and Switzerland), Australia, US or Japan. For Copilot, a policy can, according to GitHub, keep processing within the US or the EU. According to the documentation, this plan can use the hosted server at https://copilot-api.<subdomain>.ghe.com/mcp. GitHub Enterprise Server supports only the local server.

What the community offers

We found GitMCP, an open-source hosted server that, according to its README, turns GitHub projects into a documentation source. Zapier and viaSocket also offer GitHub actions through their own MCP servers. Their pages do not show how complete or current these actions are. We also found cyanheads/github-mcp-server. Its README calls the project deprecated and points to the official server. According to its README, the former reference implementation in the modelcontextprotocol repository has also moved to the official server. Claude itself also offers a GitHub integration that loads repository content into chats and projects. According to the Claude documentation, Claude reads file names and contents, but not commit history, pull requests or issues. This is not an MCP server.

What this means for your company

• The server can write. Start with read-only mode and a few toolsets.
• A token or an OAuth approval decides what the assistant sees. Grant only the rights needed, ideally per project.
• In organizations with Copilot Business or Enterprise, the “MCP servers in Copilot” policy must be enabled for Copilot to use the server. OAuth App, GitHub App and PAT policies apply to third-party hosts such as Claude.
• If you must keep data in the EU, you need GitHub Enterprise Cloud with data residency. Check the feature scope there, because according to GitHub preview features arrive only with general availability.
• What reaches the AI assistant falls under its terms, no longer under GitHub’s.