github-mcp-demo

A small, working MCP (Model Context Protocol) server that wraps the GitHub REST API as four tools an AI agent can call directly: search repositories, get repository details, list issues, and read a README.

This exists for two reasons:

1. Portfolio proof — a real, tested MCP server you can point to (GitHub pin, Fiverr/Contra gig samples, Upwork portfolio) that shows exactly what the $99 "custom MCP server" gig delivers.

2. Reusable boilerplate — the fastest way to deliver a paid job is to copy this folder and swap the API-specific pieces (see "Adapting this for a client" below), not start from a blank file each time.

What it does

Related MCP server: @cloud9-labs/mcp-github

Setup

npm install
npm run build

This produces dist/index.js, a standard stdio-based MCP server.

Optional: raise the rate limit

Without a token, GitHub allows 60 unauthenticated API requests/hour per IP, shared across anything else on that network — you'll hit this fast in testing. Set a token for real use:

export GITHUB_TOKEN=ghp_yourPersonalAccessToken

A read-only, no-scopes personal access token is enough for public repos.

Connect to Claude Desktop

Add this to your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "github-demo": {
      "command": "node",
      "args": ["/absolute/path/to/github-mcp-demo/dist/index.js"],
      "env": {
        "GITHUB_TOKEN": "ghp_yourPersonalAccessToken"
      }
    }
  }
}

Restart Claude Desktop, then try asking it things like:

• "Search GitHub for popular MCP servers written in TypeScript"

• "What are the open issues on modelcontextprotocol/typescript-sdk?"

• "Show me the README for anthropics/anthropic-sdk-python"

Adapting this for a client (the actual delivery workflow)

This is the part that makes a $99, 48-hour turnaround realistic:

1. Copy this whole folder, rename it.

2. In src/index.ts, replace GITHUB_API_BASE and the auth header logic in githubRequest() with the client's API base URL and auth scheme (API key header, Bearer token, Basic auth — same shape, different values).

3. Replace the four registerTool(...) blocks with tools matching their API's endpoints. Keep the same pattern: a Zod inputSchema for arguments, a fetch call, a small object shaping the response, textResult(...).

4. npm run build, run it through the same test pattern (spin up an MCP client over stdio, call each tool once, check the output).

5. Deliver dist/index.js + a short README with their own Claude Desktop config snippet filled in.

Steps 2–3 are the only genuinely custom work per client — everything else (project scaffold, error handling, stdio wiring, response shaping pattern) is already done.

Notes

• Errors from the upstream API (rate limits, 404s, bad auth) are caught and returned as a readable message rather than crashing the server — this matters more than it sounds like once a client is testing it themselves.

• get_readme truncates long files to ~6000 characters so a huge README doesn't eat the calling model's whole context window on one tool call.