GitHub Analyzer MCP Server

MCP (Model Context Protocol) server that exposes GitHub repository data as tools an AI assistant can call — repo info, issues, pull requests, commit statuses, PR diffs, and recent commits.

Works with any MCP-compatible client (Claude Desktop, Cline, Claude.ai, etc.) both locally (stdio) and as a deployed remote server (Streamable HTTP).

Live server: https://github-analyzer-mcp-keih.onrender.com MCP endpoint: https://github-analyzer-mcp-keih.onrender.com/mcp

What is MCP?

The Model Context Protocol is a standard way for AI applications to discover and call external tools. Instead of writing a custom integration for every AI app, you write one MCP server, and any compatible client can connect to it. This server exposes six read-only tools backed by GitHub's REST API via Octokit.

Related MCP server: Narad GitHub Agent

Tools

Tech stack

• Node.js + @modelcontextprotocol/sdk — MCP server implementation

• Octokit — GitHub API client

• Zod — input schema validation for tools

• Express — HTTP layer for Streamable HTTP (remote) transport

• Jest — unit testing

• Docker — containerized local/deployable runtime

• GitHub Actions — CI (tests + Docker build verified on every push)

• Render — deployment

Project structure

github-analyzer-mcp/
├── .github/
│   └── workflows/
│       └── ci.yml            # GitHub Actions — runs tests + Docker build on every push
├── src/
│   ├── github/
│   │   └── client.js       # Octokit setup + GitHub wrapper functions (MCP-agnostic)
│   │   └── client.test.js    # Unit tests for client.js (mocked Octokit, no live API calls)
│   ├── tools/
│   │   └── index.js        # createMcpServer() — registers all tools with schemas
│   │   └── index.test.js     # Integration tests — real MCP Client/Server over an in-memory transport
│   └── server.js            # Entry point — picks stdio or HTTP transport
├── Dockerfile                  # Container build (HTTP mode)
├── .dockerignore                # Excludes node_modules, .env, tests from the image
├── .env                      # GITHUB_TOKEN (not committed)
├── .gitignore
└── package.json

github/client.js contains plain async functions with no MCP dependency — they can be tested or reused independently of the MCP layer. Each function accepts an optional Octokit client parameter (dependency injection), so tests can pass in a fake client instead of hitting the real GitHub API. tools/index.js wraps those functions with MCP tool schemas and descriptions. server.js wires everything to a transport.

Setup

1. Clone and install

git clone https://github.com/meetvora79/github-analyzer-mcp.git
cd github-analyzer-mcp
npm install

2. Create a GitHub Personal Access Token

1. GitHub → Settings → Developer settings → Personal access tokens → Fine-grained tokens

2. Generate new token, scope it to specific repos, and grant these permissions (read-only):

   • Contents

   • Issues

   • Pull requests

   • Commit statuses

   • Metadata (usually auto-selected)

3. Configure environment variables

Create a .env file in the project root:

GITHUB_TOKEN=github_pat_xxxxxxxxxxxxxxxxxxxxxx

Running locally

stdio mode (default — for local MCP clients like Cline or Claude Desktop, which spawn the server as a subprocess):

node src/server.js

HTTP mode (Streamable HTTP — for testing the remote-capable version locally):

# macOS/Linux
MCP_TRANSPORT=http node src/server.js

# Windows PowerShell
$env:MCP_TRANSPORT="http"; node src/server.js

Server listens on http://localhost:3000, with the MCP endpoint at http://localhost:3000/mcp.

Debugging with MCP Inspector

npx @modelcontextprotocol/inspector node src/server.js

Opens a local web UI to list and manually call tools without needing a full AI client — useful for verifying a tool works before wiring up a client.

Testing

npm test

Runs the full Jest suite:

• src/github/client.test.js — unit tests for every GitHub-wrapper function (getRepoInfo, listOpenIssues, listPullRequests, getCommitStatus, getPullRequestDiff, listRecentCommits). Each function is tested against a fake Octokit client passed via dependency injection, so tests run instantly with no live GitHub API calls and no rate-limit risk.

• src/tools/index.test.js — integration tests that spin up a real McpServer and a real MCP Client connected over an in-memory transport, then call each registered tool exactly as a real MCP client would. Covers the success path, error handling (isError: true on a thrown exception), and Zod input-validation rejections for missing/invalid arguments. Because this project uses ES Modules ("type": "module"), Jest runs with Node's experimental VM modules flag — see the test script in package.json if you're curious how that's wired up.

Docker

The server can run in a container, using the same HTTP transport as the Render deployment.

Build the image:

docker build -t github-analyzer-mcp .

Run it, passing your token in at runtime (never baked into the image — .env is excluded via .dockerignore):

docker run -p 3000:3000 --env-file .env github-analyzer-mcp

The MCP endpoint is then available at http://localhost:3000/mcp.

Anyone who pulls this image gets a working server with no GitHub access until they supply their own token via -e GITHUB_TOKEN=... or --env-file. Pulling the image never grants access to the maintainer's repos or token — each user brings their own.

Connecting a client

Cline (VS Code extension)

Open Cline's MCP settings (MCP Servers icon → Configure → Configure MCP Servers) and add:

{
  "mcpServers": {
    "github-analyzer": {
      "command": "node",
      "args": ["/absolute/path/to/github-analyzer-mcp/src/server.js"],
      "env": { "GITHUB_TOKEN": "github_pat_xxxxxxxxxxxxxxxxxxxxxx" },
      "disabled": false,
      "autoApprove": []
    },
    "github-analyzer-remote": {
      "url": "https://github-analyzer-mcp-keih.onrender.com/mcp",
      "type": "streamableHttp",
      "disabled": false,
      "autoApprove": []
    }
  }
}

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "github-analyzer": {
      "command": "node",
      "args": ["/absolute/path/to/github-analyzer-mcp/src/server.js"],
      "env": { "GITHUB_TOKEN": "github_pat_xxxxxxxxxxxxxxxxxxxxxx" }
    }
  }
}

Fully restart the client after editing its config — MCP servers are only loaded on startup.

Deployment (Render)

1. Push this repo to GitHub.

2. Render dashboard → New → Web Service → connect the repo.

3. Settings:

   • Build command: npm install

   • Start command: node src/server.js

4. Environment variables:

   • GITHUB_TOKEN — your token

   • MCP_TRANSPORT — http

5. Deploy. Your MCP endpoint will be https://<your-service>.onrender.com/mcp.

Free-tier Render instances spin down after inactivity — the first request after idle time may take 30–60 seconds while the instance wakes up.

Continuous Integration

Every push and pull request to main triggers a GitHub Actions workflow (.github/workflows/ci.yml) with two jobs, run in parallel:

• test — installs dependencies and runs the full Jest suite (see Testing above).

• docker-build — runs docker build against the Dockerfile to confirm the image still builds cleanly, catching a broken Dockerfile before it reaches a real deploy. Both must pass for the workflow to go green. See the badge at the top of this README, or the Actions tab for run history.

Transport notes

• stdio — client spawns the server as a local subprocess, communicates over stdin/stdout. No networking; can't be used remotely.

• Streamable HTTP — server runs as a persistent HTTP service; clients connect over a URL. This server runs in stateless mode (a fresh server + transport instance per request) since all tools are independent, read-only lookups with no need for session memory.

Security notes

• The GitHub token lives server-side only — clients never see it directly.

• This deployment has no auth on the /mcp endpoint; anyone with the URL can call it against your token's quota. Fine for a learning project, but add authentication before using this with a real/high-limit token.

• Large payloads (e.g. PR diffs) are truncated to avoid flooding a model's context window.

Author

Meet Vora

• Email: meetvora877@gmail.com

• LinkedIn: https://linkedin.com/in/meetvora79