mcp-starter-kit
# Production-Ready MCP Server Boilerplate
> **Ship reliable MCP servers in minutes, not days.**
> A production-ready foundation for building [Model Context Protocol](https://modelcontextprotocol.io/) servers with built-in security, authentication, and observability.
> [!NOTE]
> This is a starter template β not a finished product. You are responsible for reviewing and securing any code before production use.
[](https://www.typescriptlang.org/)
[](LICENSE)
[](https://modelcontextprotocol.io/)
**Don't build from scratch.** This kit solves the "boring" parts of MCP server developmentβtransport management, error handling, and security boundariesβso you can focus on your tool's logic.
Works with **Claude Code**, **Cursor**, **Windsurf**, and any MCP-compatible client.
---
## π Why this boilerplate?
Most MCP tutorials show you a "toy" server. This repo gives you a **production-grade** foundation.
### π‘οΈ Secure by Design
- **SSRF Protection:** Blocks requests to private/internal networks (IPv4 + IPv6), including cloud metadata endpoints (`169.254.169.254`). DNS resolution is checked to prevent rebinding.
- **JWT Security:** Algorithm validation rejects `alg:none`, HS384, and RS256 downgrade attacks. Expiry and tampering are verified with constant-time comparison.
- **HMAC-SHA256 Webhook Signatures:** Outbound webhooks are signed with `X-Webhook-Signature` when a secret is configured.
- **DNS Rebinding Protection:** HTTP transport validates `Host` header against an allowlist.
- **Sandboxed File Access:** Prevents AI from reading/writing outside allowed directories, with symlink escape detection.
- **Strict Input Validation:** All tool inputs are validated with **Zod** schemas.
- **Injection Protection:** SQLite queries use strict parameter binding.
- **Security tested:** 30+ security-focused test cases covering OWASP top threats (SSRF, injection, path traversal, auth bypass).
### π Built for Real Projects
- **Authentication:** Built-in strategies for **API Key** and **JWT** (configurable per env).
- **Rate Limiting:** Token bucket algorithm to prevent abuse.
- **Observability:** Structured JSON logging (via `pino`) ready for CloudWatch/Datadog.
### β‘ Developer Experience (DX)
- **Type-Safe:** `strict: true` TypeScript configuration, ESM, fully typed.
- **Testing:** 228 tests (unit, integration, and E2E) with **Vitest**, including 30+ security-focused cases.
- **Dockerized:** Multi-stage Dockerfile for immediate deployment.
---
## π¦ What's Included
| Feature | This Boilerplate | Basic Tutorials |
| :--- | :--- | :--- |
| **Transport** | **HTTP (SSE) + Stdio** | Stdio only |
| **Validation** | **Zod Schemas** | Manual / None |
| **Logging** | **Structured JSON** | `console.log` |
| **Error Handling** | **Graceful + MCP Codes** | Process crash |
| **CI/CD** | **GitHub Actions** | None |
### Reference Implementations
Includes 6 fully-typed tools to copy-paste patterns from:
1. **`database-query`**: Secure SQLite operations.
2. **`api-connector`**: Fetch data from external REST APIs.
3. **`file-manager`**: Safe file system operations (5 sub-tools).
4. **`cache-store`**: TTL-based key-value cache with namespaces (5 sub-tools).
5. **`semantic-search`**: Local RAG with embeddings (3 tools + 2 resources + 1 prompt).
6. **`webhook-notifier`**: Async webhook delivery with task tracking (4 tools + 2 resources + 1 prompt).
---
## π Quick Start
### Prerequisites
- Node.js 22+
> **WSL2 users:** Use Node.js installed inside WSL, not Windows. See [Troubleshooting](docs/TROUBLESHOOTING.md#13-wsl2-npm-install-fails-with-eperm).
- npm 9+
### 1. Setup
```bash
git clone <your-repo-url> my-mcp-server
cd my-mcp-server
npm install
```
### 2. Configure Environment
```bash
cp .env.example .env
# Edit .env if needed β defaults work out of the box
```
### 3. Seed Sample Database
```bash
npm run db:seed # Populates local SQLite for testing
```
> **Note:** If the server is already running, restart it after seeding to pick up the new database.
### 4. Build
```bash
npm run build # Compiles TypeScript to dist/
```
### 5. Run Development Server
```bash
npm run dev
# Starts server in hot-reload mode
```
### 6. Verify with Inspector
```bash
npm run inspector
# Opens interactive debugger in your browser (uses dist/index.js)
```
---
## π§ Connecting to Clients
### Claude Code
Copy the example config to your project root:
```bash
cp .mcp.json.example .mcp.json
# Edit paths in .mcp.json to match your setup
```
Or add manually to your `.mcp.json`:
```json
{
"mcpServers": {
"mcp-starter-kit": {
"command": "node",
"args": ["dist/index.js"],
"env": {
"LOG_LEVEL": "info",
"DB_PATH": "./data/sample.db",
"SANDBOX_ROOT": "./data/sandbox"
}
}
}
}
```
### Claude Desktop
Add to your `claude_desktop_config.json`:
```json
{
"mcpServers": {
"my-production-server": {
"command": "node",
"args": ["/absolute/path/to/my-mcp-server/dist/index.js"]
}
}
}
```
### Cursor, Windsurf, & Others
See [docs/SETUP.md](docs/SETUP.md) for detailed connection guides.
---
## π Project Architecture
Designed for scalability:
```
mcp-starter-kit/
βββ src/
β βββ index.ts # Entry point
β βββ server.ts # Central MCP registry
β βββ config/env.ts # Environment variable validation (Zod)
β βββ tools/
β β βββ database-query/ # SQLite CRUD tool
β β βββ api-connector/ # REST API tool
β β βββ file-manager/ # Sandboxed file operations tool
β β βββ cache-store/ # TTL-based key-value cache
β β βββ semantic-search/ # Local RAG with embeddings
β β βββ webhook-notifier/ # Async webhook delivery
β βββ lib/ # Shared utilities (Logger, Guardrails)
β βββ middleware/ # Auth & rate limiting
β βββ transport/ # stdio and HTTP transports
βββ tests/ # Integration tests & helpers
βββ scripts/ # CLI tools (db:seed, create-tool)
βββ docs/ # Detailed documentation
βββ docker/ # Dockerfile & docker-compose
βββ data/ # Sample DB & sandbox directory
```
---
## π Documentation
| Document | Description |
|----------|-------------|
| [SETUP.md](docs/SETUP.md) | Detailed setup, environment variables, client configuration |
| [CUSTOMIZATION.md](docs/CUSTOMIZATION.md) | Step-by-step guide to creating new tools |
| [DEPLOYMENT.md](docs/DEPLOYMENT.md) | Deploy via npm, Docker, or cloud services |
| [ARCHITECTURE.md](docs/ARCHITECTURE.md) | Design decisions, security model, and data flow |
| [TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md) | Common errors and solutions |
| [TESTING.md](docs/TESTING.md) | Test suite organization, writing tests, security checklist |
| [SECURITY.md](SECURITY.md) | Security policy, vulnerability reporting, feature inventory |
---
## π οΈ Scripts
| Command | Description |
|---------|-------------|
| `npm run dev` | Start with hot-reload (tsx watch) |
| `npm run build` | Build to `dist/` (tsup) |
| `npm start` | Run built server (stdio) |
| `npm run start:http` | Run built server (HTTP) |
| `npm test` | Run all tests |
| `npm run lint` | Lint with ESLint |
| `npm run typecheck` | Type-check with tsc |
| `npm run inspector` | Open MCP Inspector |
| `npm run db:seed` | Seed sample SQLite database |
| `npm run create-tool` | Scaffold a new tool |
| `npm run test:coverage` | Run tests with coverage report |
---
## License
MIT Β© 2026 Edge Craft Studio
Not affiliated with, endorsed by, or certified by Anthropic or the Agentic AI Foundation.
TDQS
Scored across 15 tools
Each tool targets a distinct resource and action, with clear boundaries between file, database, cache, semantic, webhook, and task operations. The only minor overlap is task_status handling both get and update, but the description clearly separates these behaviors, so no two tools appear to do the same thing.
Naming conventions are mixed: some tools use hyphens (database-query, api-connector), others use underscores (file_read, task_status), and one is a single word (cache). The word order is also inconsistent, with most tools using noun_verb (file_read) but api-connector using a noun-noun pattern, and no uniform separator across the set.
With 15 tools, the server sits at the upper boundary of a reasonable scope for a starter kit. Each tool covers a distinct capability, but the sheer variety (files, database, cache, semantic search, webhooks, tasks) makes it feel slightly broad, yet still justifiable for its purpose.
The tool surface covers core operations for each domain, but there are notable gaps: no file_delete or file_rename, no webhook_unregister, and no task_cancel or task_delete. These are common lifecycle operations that agents would likely need, making the set somewhat incomplete.