Remote or local: two ways to run the same server
GitHub publishes one open source server in two forms. Which one you pick decides who maintains it and how people sign in.
| Remote (hosted by GitHub) | Local | |
|---|---|---|
| Where it runs | https://api.githubcopilot.com/mcp/ |
On each laptop, as Docker image ghcr.io/github/github-mcp-server or a binary built from source |
| Sign-in | OAuth in hosts that support it (VS Code 1.101+), otherwise a token | Browser OAuth on first use (token kept in memory only) or GITHUB_PERSONAL_ACCESS_TOKEN |
| Choosing tools | URL path /x/{toolset} or header X-MCP-Toolsets |
Flag --toolsets or env var GITHUB_TOOLSETS |
| Read-only | Path /readonly or header X-MCP-Readonly |
Flag --read-only |
| GitHub Enterprise | ghe.com yes, GitHub Enterprise Server no | Both, with --gh-host / GITHUB_HOST |
| Extra toolsets | copilot_spaces, github_support_docs_search |
Not available |
For most teams the remote server is the simpler choice: nothing to install, nothing to update, and OAuth means nobody pastes a long-lived token into a config file. The local server makes sense if you run GitHub Enterprise Server or want full control over the binary.
According to the README it works with VS Code, Visual Studio, Claude, Cursor, Windsurf, Zed, Codex, OpenCode, Copilot CLI and Copilot in JetBrains, Eclipse and Xcode.
Toolsets: give the AI only what the job needs
The server groups its tools into toolsets. Without settings you get the default five: context, repos, issues, pull_requests and users. Others include actions, code_security, discussions, gists, git, labels, projects, secret_protection and security_advisories. all turns everything on.
Every enabled tool takes room in the model’s context, and every write tool is something the AI can do on your behalf. A sensible setup for a team:
- Start from the defaults, or narrower:
repos,issues,pull_requests. - Add
actionsonly for people who debug CI with the AI. - Use read-only mode for anyone who mainly asks questions about the code.
- Keep write access for agents that are meant to open pull requests, and review those PRs as you would a colleague’s.
On the remote server, a header can also turn on lockdown mode (X-MCP-Lockdown), which filters public repository content based on whether its author has push access. GitHub’s own docs call that “a best-effort content filter, not a security boundary”, so it does not replace limiting what the agent may do.
Setting it up for a team, step by step
- Copilot users: ask an admin to enable the policy MCP servers in Copilot for your organization or enterprise. Without it, members on Copilot Business or Enterprise cannot use MCP servers.
- Pick remote or local per the table above.
- Add the server in each tool. In VS Code you add an HTTP server with the remote URL; in Claude Code you can store it in the project’s
.mcp.jsonso everyone who clones the repo gets the same setup (see MCP servers in Claude Code for a team). Cursor and Codex have their own config files. - Sign in per person. Each person authenticates as themselves, so the AI can only see what that person can see on GitHub. If you use a token, use a fine-grained token with the minimum scopes.
- Agree on toolsets and read-only and write them down where the team can find them.
The common misunderstanding
People expect the GitHub MCP server to make the AI “know the codebase”. It does not. It gives the AI tools to fetch files, search code, read issues and open PRs when it decides to. It does not hold knowledge of why things are built the way they are, which conventions your team agreed on, or what a reviewer corrected last week. That knowledge lives in heads, PR comments and chat threads, and the AI only sees it if it happens to search for exactly that.
A second point: the server acts with the permissions of the person who signed in. If an agent works with a token that can write to every repo, it can write to every repo. See is MCP safe to use in a company for the wider risks.
What this means for a team
The GitHub MCP server is a good shared connection: everyone’s AI reaches the same repos in the same way, whatever tool they use. What it does not share is what each person’s AI learned in its own sessions. The developer in Cursor and the developer in Claude Code still each explain the same conventions to their own agent. For how teams handle that, read shared memory across coding agents. For the general picture of what an MCP server is, see what is an MCP server.
Where Bombos fits
The GitHub MCP server gives every AI tool access to your repos; Bombos gives them the same memory about how your team works with those repos. It is a plugin for Claude Code, Codex, Cursor, Copilot and the other AI tools a team uses, holding team knowledge and rules the AI learned from corrections, and it reads what already sits in repos, documents and mail, with a trail of what each answer rests on. It runs in our cloud or on your own server, and one developer can try it first with their own agents.
Related questions
Is the GitHub MCP server free?
The server itself is open source on GitHub. You need a GitHub account, and the AI tool you connect it to may have its own price.
Do I need Docker to use it?
Only for the local version, which runs as the container image ghcr.io/github/github-mcp-server. The remote version needs no installation, just an MCP host that supports remote servers.
Does it work with GitHub Enterprise Server?
The local server does, with the --gh-host flag or GITHUB_HOST pointing to your https host. The remote hosted server supports ghe.com but not GitHub Enterprise Server.
Why can't Copilot in my company use MCP?
On Copilot Business and Enterprise an admin must enable the policy "MCP servers in Copilot" for the organization or enterprise before members can use MCP servers.
Can I stop the AI from merging or pushing?
Yes. Start the local server with --read-only, or send the X-MCP-Readonly header (or use a /readonly URL path) on the remote server. Read-only wins even if a write tool is named explicitly.
