GitHub MCP Server

The GitHub MCP server is the one most developers install first, and the official version (built by GitHub, in Go) has grown into a genuinely deep integration: repository browsing, file contents, issue and PR management, code and issue search across all of GitHub, Actions workflow runs, code-scanning alerts, Dependabot findings, notifications and discussions.

The workflow change is bigger than it sounds. Instead of alt-tabbing to github.com to check CI, you ask "why did the last workflow run fail on main?" and get the failing step's logs interpreted in context. PR reviews become conversational: the assistant reads the diff, summarises the risk areas, drafts review comments, and can request changes or merge when you say so. Issue triage — labelling, deduplicating, drafting responses — is the kind of tedium this quietly absorbs.

Two ways to run it. The remote server at api.githubcopilot.com/mcp/ is the low-friction option: OAuth in the browser, no local install, always current. The local version runs as a Docker container with a personal access token — the right choice for GitHub Enterprise Server or when you want tight control over scopes via a fine-grained PAT.

A practical tip: the server supports toolsets — flags that trim which tool groups are exposed (just repos and PRs, say). Trimming unused toolsets reduces prompt clutter and keeps the assistant from reaching for tools you never intended it to use. Scope your PAT minimally and this becomes a very comfortable daily driver.

Tuning the toolset surface

• Set GITHUB_TOOLSETS=repos,issues,pull_requests to expose only what you need; all turns everything on, and dynamic toolset discovery lets the model switch groups on mid-session instead of loading them all upfront.
• The remote endpoint has per-toolset paths such as /mcp/x/issues and /mcp/x/actions, plus a read-only variant — so you can register a browsing-only GitHub server next to a write-capable one.
• Fine-grained PAT permissions have to match the toolsets you enable: Contents for file reads, Pull requests for review flows, and Actions read access before job logs will come back at all.
• get_job_logs can return only the failed steps with a tail limit. Reach for that rather than pulling a 40,000-line workflow log into the conversation.