Copilot in VS Code can read the files you have open. Out of the box it cannot read the issue you are fixing, the review comments on your pull request, or the log of the GitHub Actions run that just failed. You end up copying all of that into chat by hand.
The GitHub MCP Server is GitHub's official bridge between AI tools and the GitHub platform. Once connected, your agent can browse code across repos you can access, create and update issues and pull requests, look at workflow runs and build failures, and read security findings and Dependabot alerts, all from a chat prompt.
GitHub hosts a remote version, so in VS Code there is nothing to install or run. This guide covers the setup, how to limit what the agent can do, and a local option if you want one.
The README has an "Install in VS Code" button. Click it and complete the flow, then toggle Agent mode next to the Copilot Chat input and the server starts. VS Code 1.101+ supports OAuth for remote servers, so you can sign in with your GitHub account instead of creating a token.
Open (or create) .vscode/mcp.json in your project and add the hosted server:
{
"servers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/"
}
}
}
That uses OAuth. If you would rather use a GitHub personal access token, the README's version prompts you for it as a password input instead of hard-coding it in the file:
{
"servers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"headers": {
"Authorization": "Bearer ${input:github_mcp_pat}"
}
}
},
"inputs": [
{
"type": "promptString",
"id": "github_mcp_pat",
"description": "GitHub Personal Access Token",
"password": true
}
]
}
Switch Copilot Chat to Agent mode and ask for things that used to take a tab switch:
List the open issues labeled bug in this repo and summarize the three oldest.
Read the review comments on my open pull request and fix what they ask for.
Why did the last GitHub Actions run on main fail? Show me the failing step.
By default the server turns on five toolsets: context, repos, issues, pull_requests and users. GitHub's own advice is that enabling only what you need helps the model pick the right tool and keeps the context smaller. Other toolsets include actions, code_security, dependabot, discussions, notifications, projects and more.
With the remote server you control this through the URL or headers.
Read-only. Add /readonly to the URL and the agent gets only read tools. A good default while you build trust:
{
"type": "http",
"url": "https://api.githubcopilot.com/mcp/readonly"
}
One toolset. Use /x/{toolset}, for example issues only, read-only:
{
"type": "http",
"url": "https://api.githubcopilot.com/mcp/x/issues/readonly"
}
Several toolsets. The path takes one toolset at a time, so for a mix use the X-MCP-Toolsets header:
{
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"headers": {
"X-MCP-Toolsets": "repos,issues,actions",
"X-MCP-Readonly": "true"
}
}
/x/all turns on everything, and /insiders (or the X-MCP-Insiders header) opts you into experimental tools early.
If your agent reads issues and pull requests on a public repo, anyone can write text it will see. The server has a lockdown mode that hides content from authors without push access. Turn it on with the X-MCP-Lockdown: true header. GitHub is clear that this is a best-effort filter, not a security boundary: it does not change what your token can read or write. Pair it with read-only mode and a token scoped to the minimum.
You can run the server in Docker instead. On github.com the one-click Docker buttons log you in with OAuth on first use. With a personal access token, the README's Docker form is:
docker run -i --rm \
-e GITHUB_PERSONAL_ACCESS_TOKEN=<your-token> \
-e GITHUB_TOOLSETS="repos,issues,pull_requests,actions,code_security" \
ghcr.io/github/github-mcp-server
Add -e GITHUB_READ_ONLY=1 for read-only mode. The same image works with Claude Code:
claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=$GITHUB_PAT -- docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN ghcr.io/github/github-mcp-server
The repo also has install guides for Cursor, Claude Desktop, Codex, Windsurf, Zed and Copilot in JetBrains, Visual Studio, Eclipse and Xcode.
From the README's own checklist: grant only the scopes you need, use separate tokens per project, rotate them, never commit them (keep them in an environment variable or a gitignored .env), and restrict permissions on config files that contain them.
Our short write-up of this setup, with who it fits and what to watch for, is here: How to connect your agent to GitHub with the hosted MCP server. More ways people use it are on the GitHub MCP Server page on Deduped.
GitHub is usually the first MCP server people add, not the last. Our ranked list of the best open-source MCP servers covers what to add next.
Deduped tracks the AI repos people are actually building with and writes up how to use them. Repo: github.com/github/github-mcp-server.
Get one email every Monday with the week's hottest AI repos and exactly how to use them: subscribe free at dedupedai.com