github-mcp
Self-hosted GitHub MCP server. Personal-access-token auth, stdio transport, transport-agnostic so it works with Claude Code (any account), Codex CLI, Cursor, or any MCP-compliant client. No claude.ai OAuth dependency.
Written in Python on top of FastMCP. Slots into the same install pattern as other community Python MCP servers.
Why this exists
The official github/github-mcp-server (Go, 60+ tools) is excellent if you want a kitchen-sink server tied to a single agent platform's auth flow. This one is scoped for operators who:
- Switch between multiple Claude accounts, Codex CLI, Cursor, or other MCP clients and want the same GitHub auth across all of them.
- Prefer a small, audited tool surface scoped to PR + repo file + Dependabot operations.
- Want a Python install that lives alongside their other MCP servers.
If neither applies, run github/github-mcp-server instead โ that's the right call.
Tool surface (v0.1.0)
| Tool | Purpose |
|---|---|
health_check |
Verify the PAT is valid + report rate-limit headroom. First call when wiring up. |
list_prs |
List pull requests with state filter. |
get_pr |
Single PR with check-run status + mergeable state. |
merge_pr |
Squash/merge/rebase with optional branch delete. |
list_dependabot_alerts |
Alert sweep with severity-grouped counts. |
get_file |
Read a file via the Contents API. |
put_file |
Create or update a file via the Contents API (idempotent with SHA). |
Future versions expand: workflows, releases, issues, rulesets, secret-scan alerts, branch protection, repo settings.
Install
git clone https://github.com/adelaidasofia/github-mcp ~/.claude/github-mcp
cd ~/.claude/github-mcp
uv tool install --editable .
Or as a one-off without persistent install:
cd ~/.claude/github-mcp
uv run github-mcp
Auth
The server reads the GitHub PAT from environment, in this precedence:
GITHUB_TOKENGH_TOKEN
Fine-grained tokens recommended.
Scope matrix
| Permission | Access | Unlocks |
|---|---|---|
| Metadata | Read | Baseline (auto-required when any other repo permission is set) |
| Contents | Read and write | get_file, put_file |
| Pull requests | Read and write | list_prs, get_pr, merge_pr |
| Dependabot alerts | Read | list_dependabot_alerts |
| Administration | Read and write | Future: rulesets, branch protection, allow_auto_merge toggle |
| Workflows | Read and write | Future: read/update .github/workflows/*.yml |
| Actions | Read and write | Future: trigger and inspect workflow runs |
| Secret scanning alerts | Read | Future: fleet-wide secret-scan sweep |
| Code scanning alerts | Read | Future: CodeQL fleet sweep |
v0.1 floor: Metadata + Contents (R/W) + Pull requests (R/W) + Dependabot alerts (R) is enough to use every tool in this release.
Future-proof: granting the full table now means new MCP tools can ship without rotating the PAT. Equivalent to checking "all repository permissions" in the fine-grained PAT UI. Trade-off: broader blast radius if the token leaks. Reasonable for a personal-dev token on an encrypted laptop with FileVault; reconsider for shared or production deployments.
The MCP introspects scopes via health_check โ call it after wiring up to confirm what your token actually has.
Optional: GITHUB_API_BASE for GitHub Enterprise Server. Defaults to https://api.github.com.
Wire into an MCP client
Claude Code
Add to ~/.claude/.mcp.json (or any .mcp.json in the project root):
{
"mcpServers": {
"github": {
"command": "uv",
"args": ["run", "--project", "/Users/<you>/.claude/github-mcp", "github-mcp"],
"env": {
"GITHUB_TOKEN": "${GITHUB_TOKEN}"
}
}
}
}
Codex CLI
Add to ~/.codex/config.toml:
[mcp_servers.github]
command = "uv"
args = ["run", "--project", "/Users/<you>/.claude/github-mcp", "github-mcp"]
[mcp_servers.github.env]
GITHUB_TOKEN = "${GITHUB_TOKEN}"
Cursor
~/.cursor/mcp.json follows the Claude Code shape. Same env block.
Any other MCP-compliant client
Configure it to spawn the github-mcp command and pipe stdio. The server speaks standard MCP over stdio.
Verify
GITHUB_TOKEN=ghp_yourpat uv run github-mcp
The server will wait on stdin for MCP traffic. In another shell, exercise health_check via your client.
Test
uv pip install -e ".[dev]"
uv run pytest tests/ -v
License
MIT.