GitHub MCP Server
Issues, pull requests and code search without leaving the agent
GitHub's official MCP server gives an agent typed access to repositories, issues, pull requests, actions and code search. It replaces a pile of gh CLI calls with a defined tool surface and a scoped token.
What it changes
Reviewing a pull request, triaging issues or searching across an organisation becomes something the agent does directly rather than something you paste into it. Code search across repositories is the underrated part, because it answers questions a single checkout cannot.
Because access is a scoped token rather than your whole shell, you can give an agent read access to an organisation without giving it the ability to push.
Scoping it properly
Start read only. An agent that can open pull requests and edit issues is useful and also writes to a place other people are watching, so earn that permission deliberately.
Large repositories can return very large responses. Ask for specific files and specific pull requests rather than whole listings, or the context fills with material nobody reads.
Reach for it when
- triaging issues or reviewing pull requests with an agent
- searching code across many repositories at once
- giving an agent scoped repository access instead of your shell
Watch out for
- write access means the agent can post where colleagues will see it, so start read only
- listings on large repositories return a lot of data, so ask narrowly
- the token you issue is the real security boundary, so scope it rather than trusting prompts
Compare with these
- MCP serversFree
Chrome DevTools MCP
Let the agent open the page and read what actually rendered
- mcp
- browser
- chrome
Apache-2.0Read - MCP serversFree
Playwright MCP
Browser automation through the accessibility tree instead of pixels
- mcp
- browser
- playwright
Apache-2.0Read - MCP serversFree tier
Context7
Current library documentation, fetched instead of remembered
- mcp
- documentation
- context
MITRead