
github-mcp-server
GitHub's own MCP server: issues, pull requests and Actions over the API, never a local checkout
What is github-mcp-server?
Point Claude, Copilot, Cursor or any other MCP client at it and the model can triage an issue, open and review a pull request, search code it has never downloaded, and read the logs of a failed Actions run — all through the GitHub API, so nothing is cloned. The remote server GitHub hosts needs no local setup or runtime at all; the same local server runs as a container or a single binary, and is the only route for GitHub Enterprise Server. The trade-off is breadth: more than eighty tools sit across twenty-odd toolsets, and only five toolsets are on by default for good reason — turning everything on gives most models more choices than they handle well.
What can you do with github-mcp-server?
- Take an issue through to a merged pull request — It reads an issue with its comments, its parent and its sub-issues, sets labels, assignees and milestone, opens a pull request, leaves review comments and merges. It can also create a branch and push several files in one commit, so a one-line fix never needs a working copy.
- Search code and history without cloning — Code search, commit search and file reads work on any repository the credential can see, which makes it practical to ask about a codebase you have never downloaded. Asking for the repository tree first stops the model guessing at file paths.
- Follow a red build down to the failing log — It lists workflows, their runs and the jobs inside a run, and can fetch the logs of only the failed jobs in one run, trimmed to the last few lines you ask for. A CI failure then arrives at a size that fits in the conversation instead of flooding it.
- Switch off the tools you are not using — Five of the twenty-odd toolsets are enabled by default; you name the ones you want, add or exclude single tools on top, and make the same choices on the remote server through the URL path or a request header. Fewer tools means sharper tool choice and less context spent.
- The credential decides which tools even appear — With a classic personal access token the server reads the token's permissions at startup and hides the tools it could not use anyway. Tokens that do not publish their permissions — the newer repository-scoped ones, and GitHub App tokens — get the full list instead, and a call you lack permission for fails at the API. On the remote server, signing in through the browser prompts you to approve an extra permission the moment a tool needs it.
- Read-only mode, and lockdown mode for public repositories — Read-only strips every write tool and overrides anything else you asked for, so an agent can look but never touch. Lockdown mode goes further on public repositories: the body of an issue or pull request written by someone without push access comes back as an error, and their comments and reviews are filtered out, so a stranger's text cannot quietly instruct your agent.
Before you choose github-mcp-server
- The server has no repository setting, so an agent reaches wherever its credential does; confining it to one project is done in the credential — a token scoped to those repositories, or a GitHub App installed only there.
- Images attached to issues and pull requests in private repositories come back as links that need a signed-in browser session, and as of August 2026 the server has no tool that fetches them.Reported in#807
Star history
17 Aug to 30 Oct · +1.8k
Frequently asked questions
Is github-mcp-server free for commercial use?
github-mcp-server is released under the MIT licence — OSI-approved open source, which permits commercial use.
How can github-mcp-server be deployed?
github-mcp-server is available as Runs locally / Managed cloud / Self-hosted.