GitHub MCP Server

The GitHub MCP Server is an open-source Go server published by GitHub (repo: github/github-mcp-server) that exposes the GitHub REST + GraphQL APIs as MCP tools. Your AI assistant — Claude Code, Cursor, Claude Desktop, VS Code Copilot, Windsurf, JetBrains — gets a uniform JSON interface to do anything you can do with the gh CLI: open issues, review diffs, fetch CI logs, scan vulnerabilities.

All configs below assume the remote server. For Docker stdio, swap the type: "http" block for command/args as shown at the bottom.

Claude Code (CLI)

One command — Claude Code handles OAuth in the browser:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/

Cursor

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/"
    }
  }
}

Claude Desktop

Edit claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "mcp-remote@latest", "https://api.githubcopilot.com/mcp/"]
    }
  }
}

VS Code (with Copilot)

VS Code 1.101+ supports MCP natively. Add to .vscode/mcp.json:

{
  "servers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/"
    }
  }
}

Docker stdio (Enterprise / offline)

{
  "mcpServers": {
    "github": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "GITHUB_PERSONAL_ACCESS_TOKEN",
        "-e", "GITHUB_TOOLSETS=repos,issues,pull_requests",
        "-e", "GITHUB_READ_ONLY=1",
        "ghcr.io/github/github-mcp-server"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx..."
      }
    }
  }
}

☝ Note: GITHUB_READ_ONLY=1 disables every write tool. Use this for QA assistants by default.

Toolsets group related tools so you can enable only what you need. Smaller context = better tool choice = fewer hallucinations.

Set via flag: --toolsets=repos,issues,pull_requests or env: GITHUB_TOOLSETS=.... Enable experimental tools with GITHUB_INSIDERS=true or URL /mcp/insiders.

The biggest mistake QA teams make is plugging in a PAT with repo scope and forgetting that merge_pull_request is in the toolset. The AI will use it if asked. Use this matrix to decide what to enable.

Each prompt is tested with Claude Code + Cursor against a real repo. Replace {OWNER}/{REPO} with your repo path.

Use the GitHub MCP to list open pull_requests against the default branch from the last 7 days for repo {OWNER}/{REPO}.

For each PR:
1. Fetch the diff (get_pull_request_files).
2. Identify if any changed file matches these high-risk paths: src/payment/**, src/auth/**, src/migrations/**.
3. List the PR number, title, author, files changed, and a 1-sentence risk note.

Format as a markdown table sorted by risk (high → low). Do not comment on or merge any PR.

Using GitHub MCP (toolset: actions), list the 20 most recent failed workflow runs for repo {OWNER}/{REPO} on the main branch.

For each failure:
1. Fetch the failing job logs (get_job_logs).
2. Extract the test name(s) that failed and the assertion message.
3. Group failures by test name.
4. For any test that failed >2 times across different commits, mark it as "flaky candidate" and propose a hypothesis (timing, network, shared state).

Output: markdown report with sections "Flaky candidates" and "One-off failures".

Using GitHub MCP, find all PRs merged into main between the last release tag and HEAD for repo {OWNER}/{REPO}.

For each merged PR:
1. Read the PR title, body, and labels.
2. Categorize: "Features" (label: enhancement), "Fixes" (label: bug), "Other".
3. Skip PRs labeled "internal" or "ci".
4. Rewrite each item in a 1-line, user-facing tone (no jargon, no PR numbers in the line itself).

Output: markdown grouped by category, ready to paste into release notes.