ghe-mcp-gateway

An MCP (Model Context Protocol) server that exposes GitHub Enterprise administration operations as agent tools. Point Claude (Code or Desktop) at it and ask things like "list private repos in the org with no recent pushes" or "who has admin on payments-api?" - the model calls the GitHub API through typed, permissioned tools instead of raw curl.

Built as a working reference for GitHub Enterprise Cloud administration: repository management, access governance, audit-log review, webhook management, and secret-scanning triage.

Why

Admin work is repetitive API calls (access reviews, alert triage, config audits). Wrapping the GitHub REST/GraphQL API as MCP tools lets an agent do the legwork while keeping a human in the loop - and keeps writes behind an explicit safety switch.

Tools (21)

* write operation - blocked unless GITHUB_MCP_READ_ONLY=false.

Companion skill

.claude/skills/github-admin/SKILL.md packages these tools into admin playbooks - access review/certification, leaked-secret response, repo onboarding & governance, security-posture audit, and access changes. Open Claude Code in this repo and the /github-admin skill is available; it drives the MCP tools with a read-first, least-privilege, confirm-before-write discipline.

Setup

uv sync                       # create venv + install deps
cp .env.example .env          # then add your GITHUB_TOKEN

Suggested token scopes (classic): repo, read:org, admin:org, read:audit_log, admin:org_hook, security_events. For GitHub Enterprise Server, set GITHUB_API_URL to https://<host>/api/v3.

Run

# stdio server (how MCP clients launch it)
uv run github-admin-mcp

# quick manual check with the MCP Inspector
uv run mcp dev src/github_admin_mcp/server.py

Register with Claude Code

claude mcp add github-admin -- uv run --directory /Users/mikeholzinger/src/github_mcp github-admin-mcp

(or add an entry to your client's MCP config pointing at the same command).

60-second demo

# 1. add your token (read-only by default - safe)
cp .env.example .env && $EDITOR .env        # set GITHUB_TOKEN

# 2. register the server + open Claude Code in this repo
claude mcp add github-admin -- uv run --directory "$PWD" github-admin-mcp
claude

# 3. the /github-admin skill is now available. Try, in natural language:
#    "Run an access review on the <org> organization"
#    "Audit the security posture of <org>/<repo>"
#    "Who has admin on <org>/<repo>?"
# The skill calls the MCP tools read-only and reports an auditor-ready summary.

Safety

• Read-only by default (GITHUB_MCP_READ_ONLY=true); mutating tools refuse until you opt in.

• No credentials in code - token comes from the environment. .env is git-ignored.

Layout

src/github_admin_mcp/
  client.py   # async REST + GraphQL client (auth, pagination, GHES-aware, read-only guard)
  server.py   # FastMCP server; one @mcp.tool() per operation
docs/
  GITHUB_API_REFERENCE.md   # the endpoint research this server is built on

Example GitHub Actions workflows

Healthcare-oriented admin automation lives in .github/workflows/ with the logic in scripts/ (Python + requests). All follow least-privilege permissions: and pin actions.

Required CI config: secret ORG_ADMIN_TOKEN (org PAT/App), secret SECURITY_SLACK_WEBHOOK, and variable GITHUB_ORG. The scripts honor GITHUB_API_URL for GitHub Enterprise Server.

Roadmap

• Branch-protection / rulesets tools, team-repo access management

• Secret-scanning alert resolution, code-scanning + Dependabot alerts

• Audit-log streaming config; enterprise-level endpoints

• A companion Claude skill that drives these tools for common admin playbooks