GitHub MCP server: connect AI agents to GitHub from the EU

GitHub, owned by Microsoft and based in San Francisco, hosts source code, issues and CI pipelines for most software teams. Its official MCP server, github/github-mcp-server, is written in Go and published under the MIT licence.

It comes in two forms. The remote server at https://api.githubcopilot.com/mcp/ is hosted by GitHub and uses Streamable HTTP. The local server runs from a Docker image or binary over stdio, and is the option for GitHub Enterprise Server. Toolsets such as repos, issues, pull_requests, actions and code_security control which tools the agent sees.

A triage agent can read new issues each morning, label them, find duplicates and ask reporters for missing details. A review agent can summarise a pull request's diff, point out risky changes and leave a review comment for the author.

When a build fails, an agent can read the failed job's logs, explain the error and propose a fix in a new branch. Security teams can ask for open code scanning and Dependabot alerts across repositories and get a prioritised list with links.

On inference.eu, add the GitHub connector to an agent and sign in with GitHub through OAuth; the agent then acts as you, within your repository permissions. For automation accounts, you can instead paste a fine-grained personal access token limited to selected repositories and permissions.

Tokens are stored encrypted in the EU and only reach the agents you allow. Pick the toolsets each agent needs; fewer tools also help the model choose correctly. Read-only mode removes every write tool, even if a toolset includes one.

With the remote server, requests go to GitHub's own MCP endpoint and from there to the GitHub API, so your code and issues stay with GitHub, a US company, as they do today. With the local server, the process calls the GitHub API directly.

On inference.eu, the model that reads code and issues runs in EU data centres, and your OAuth token or access token is stored encrypted in the EU. Source code may contain secrets or personal data in comments and test fixtures, so scope access to the repositories an agent works on.

Public issues and pull requests can contain text written to manipulate an agent. GitHub's lockdown mode filters content from authors without push access on public repositories, which reduces that risk. Read-only mode is the safest start for analysis agents.

On inference.eu you can require a person to approve write calls, such as merging a pull request, pushing files, closing issues or rerunning workflows, while reads run freely. Use separate tokens per agent and rotate them, so access can be withdrawn without affecting other work.