derekhuynen-mcp-server

A production-grade reference MCP server in TypeScript: one core, two transports, strict types, real tests, Docker, and CI

A production-grade reference implementation of a Model Context Protocol (MCP) server in TypeScript. One transport-agnostic core, two transports (stdio + Streamable HTTP), strict types, real tests, Docker, and CI. Clone it, read it, template it.

Most MCP examples are toy, single-file stdio scripts. This is what a real MCP server looks like: a clean core that registers tools, resources, and prompts once and serves them over two transports, backed by a validated data layer, a full test suite, a multi-stage Docker build, and CI.

The data it serves happens to be a real professional profile (mine), so the example is end-to-end and honest rather than foo/bar. But the reason to star it is the architecture and the patterns, which transfer to any MCP server you want to build.

Why you might care

• The production shape of an MCP server, not a snippet: clear module boundaries, a buildServer() factory, and thin transport entrypoints.

• Two transports from one core. Run it locally over stdio (Claude Desktop / Claude Code) or deploy it as a stateless Streamable HTTP service. No logic duplicated between them.

• Actually tested. 42 tests, including a real in-memory MCP client driving the server (not mocks).

• Deploy-ready. Multi-stage Dockerfile, docker-compose, health check, CORS, rate limiting, graceful shutdown, structured logging, env-based config.

• Type-safe end to end. A Zod schema is the single source of truth, with inferred TypeScript types throughout.

• A pattern worth stealing: the served content is generated from a single source of truth into a validated, reviewed snapshot, so the public surface never drifts.

Related MCP server: personal-resume-agent

What it exposes

Tools

Resources: profile://summary, profile://resume Prompts: recruiter_pitch (arg: role)

See it in action

Once wired into an MCP client like Claude, the model can call the tools to answer questions about the data:

You:    What has Derek done with RAG?
Claude: (calls get_experience { "skill": "RAG" })
        Three RAG engagements stand out:
        - PG&E GenAI Regulatory Chatbot (Azure OpenAI + Azure AI Search)
        - HarperCollins Book Catalog RAG Chatbot (semantic search over 500 books)
        - A hospital document-processing proof of concept

The tool call behind that answer returns plain JSON:

// get_experience { "skill": "RAG" }  ->  (one entry shown, trimmed)
[
  {
    "slug": "neudesic-pge-genai-chatbot",
    "project": "GenAI Regulatory Chatbot",
    "title": "AI Developer",
    "employer": "Neudesic (an IBM Company)",
    "client": "PG&E (Pacific Gas & Electric)",
    "start": "2025-01",
    "end": "2025-03",
    "summary": "Core engineer on an enterprise RAG chatbot that let PG&E staff query regulatory, claims, and compliance documents in natural language with real-time, grounded citations.",
    "skills": ["RAG", "Azure OpenAI", "Semantic Kernel", "Azure AI Search"],
    "featured": true,
  },
]

Quick start

npm install
npm run build

Run over stdio (local MCP clients)

npm run start:stdio

Wire it into Claude Desktop (claude_desktop_config.json) or Claude Code:

{
  "mcpServers": {
    "derek": {
      "command": "node",
      "args": ["/absolute/path/to/derekhuynen-mcp-server/dist/transports/stdio.js"]
    }
  }
}

Run over HTTP

npm run start:http
# GET  http://localhost:3000/health
# POST http://localhost:3000/mcp   (Streamable HTTP MCP endpoint)

Example initialize call:

curl -X POST http://localhost:3000/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"curl","version":"1.0"}}}'

Run with Docker

docker compose up --build
curl http://localhost:3000/health

Configuration

Architecture

A transport-agnostic core (buildServer(profile)) registers all tools, resources, and prompts against an in-memory profile loaded and validated from src/data/profile.json. Two thin entrypoints connect transports to a freshly built server:

flowchart TB
    subgraph Clients["MCP clients"]
        CD["Claude Desktop / Claude Code<br/>(stdio)"]
        HC["HTTP client<br/>(Streamable HTTP)"]
    end
    subgraph Server["derekhuynen-mcp-server"]
        STDIO["transports/stdio.ts"]
        HTTPT["transports/http.ts<br/>Express: health, CORS, rate limit"]
        CORE["buildServer(profile)<br/>tools · resources · prompts"]
        Q["core/queries.ts<br/>(pure business logic)"]
    end
    DATA[("data/profile.json<br/>Zod-validated snapshot")]
    CD --> STDIO
    HC --> HTTPT
    STDIO --> CORE
    HTTPT --> CORE
    CORE --> Q
    Q --> DATA

The module layout:

src/
  data/
    schema.ts       Zod schema + inferred types (single source of truth)
    loader.ts       parse + validate the snapshot (fail fast)
    profile.json    generated, public-safe data snapshot
  core/
    queries.ts      pure business logic (unit-tested)
    tools.ts        registerTools()      thin adapters over queries
    resources.ts    registerResources()
    prompts.ts      registerPrompts()
    server.ts       buildServer(profile) factory
  transports/
    stdio.ts        stdio entrypoint
    http.ts         Express + Streamable HTTP (health, CORS, rate limit, shutdown)

The key property: one server definition, two entrypoints. Both transports call the same buildServer(), so they can never drift in what they expose.

Use it as a template

Building your own MCP server? Here is the map:

1. Replace src/data/ with your own data shape and Zod schema.

2. Rewrite src/core/queries.ts with your domain logic (keep it pure; it stays easy to test).

3. Adjust the registrations in src/core/{tools,resources,prompts}.ts.

4. Leave server.ts and the transports mostly as-is. That plumbing is the reusable part.

How the data stays honest

src/data/profile.json is not hand-written. It is generated from a separate source of truth by scripts/generate-profile.mjs (npm run generate), which strips anything private and validates the result against the schema before it can ship. This keeps the public surface accurate and is a pattern worth reusing whenever a server exposes curated data.

Development

npm run dev:http     # hot TS run of the HTTP server
npm run dev:stdio    # hot TS run of the stdio server
npm test             # vitest
npm run lint
npm run typecheck

Deploy

The repo ships deploy-ready (Dockerfile + compose). The HTTP image is stateless and runs on any container host (Azure Container Apps, Fly.io, Cloud Run, a VPS, and so on). There is no authentication because all data is public by design; the endpoint is protected by Helmet security headers, rate limiting, and CORS. Behind a proxy, set TRUST_PROXY so client IPs (and therefore rate limiting) are accurate.

Tech stack

TypeScript (strict, ESM) - @modelcontextprotocol/sdk - Zod - Express - pino - Vitest - Docker - GitHub Actions.

License

MIT (c) Derek Huynen

If this helped you understand or build an MCP server, a star is appreciated. It helps other developers find it.