Example MCP Server

mctx — The best way to Build an MCP Server

The comprehensive reference implementation for @mctx-ai/mcp. Every framework capability in one well-commented file — clone it, study it, fork it as a template for your own MCP server.

What This Example Demonstrates

src/index.ts covers every pattern the framework supports:

Tool Patterns

• Sync string return — greet: receives args, returns a formatted string; demonstrates environment variable configuration via GREETING
• Object return — calculate: returns a structured result object (auto-serialized to JSON); demonstrates input validation and error throwing
• Progress notifications — analyze: emits progress via res.progress(current, total); demonstrates streaming updates during long-running operations
• LLM sampling — smart-answer: delegates to the client's LLM via res.ask; demonstrates graceful fallback when sampling is unavailable
• User identity via context — whoami: reads mctx.userId, the stable mctx user ID injected server-side by the platform; demonstrates graceful degradation outside mctx

Resource Patterns

• Static URI — docs://readme: exact URI, no parameters, returns plain text
• Dynamic URI template — user://{userId}: extracts userId from the URI path, returns a JSON profile

Prompt Patterns

• Single-message — code-review: returns a plain string that becomes one user message
• Multi-message conversation — debug: uses conversation() to build structured user/assistant dialogue

Infrastructure

• Tool annotation hints — every tool sets readOnlyHint, destructiveHint, openWorldHint, and idempotentHint with inline rationale explaining each choice
• Structured logging — log.info, log.debug, log.warning, log.error, and log.notice throughout
• Environment variable configuration — reads process.env.GREETING lazily inside the handler (not at module scope)
• Comprehensive test coverage — src/index.test.ts tests every tool, resource, and prompt via JSON-RPC 2.0 requests

Usage in Conversation

Once connected to an MCP client, try phrases like these:

Greetings and identity

• "Greet Alice" — calls greet with name: "Alice"
• "Who am I?" — calls whoami to return your stable mctx user ID
• "Greet the whole team" — calls greet in a loop for each name

Math and analysis

• "What is 6 times 7?" — calls calculate with operation: "multiply"
• "Divide 10 by 0" — triggers the division-by-zero error path
• "Analyze quantum computing" — calls analyze and streams three progress phases

Q&A and LLM sampling

• "What is the capital of France?" — calls smart-answer, which delegates to the client's LLM via ask
• "Ask anything" — routes arbitrary questions through smart-answer

Resources

• "Read the server docs" — reads docs://readme
• "Look up user 42" — reads user://42

Prompts

• "Review my code" — invokes the code-review prompt with a code snippet
• "Help me debug this stack trace" — invokes the debug prompt with an error and optional context

Example Responses

greet(name: "Alice")
→ "Hello, Alice!"

whoami()
→ "Your mctx user ID is: user_abc123. This ID is stable across all your devices and sessions."

calculate(operation: "multiply", a: 6, b: 7)
→ { "operation": "multiply", "a": 6, "b": 7, "result": 42 }

analyze(topic: "quantum computing")
→ [progress: 1/3] → [progress: 2/3] → [progress: 3/3]
→ "Analysis of "quantum computing" complete. Found 42 insights across 7 categories."

smart-answer(question: "What is the capital of France?")
→ "Question: What is the capital of France?\n\nAnswer: Paris."

Read URI: docs://readme
→ "Welcome to the example MCP server built with @mctx-ai/mcp..."

Read URI: user://42
→ { "id": "42", "name": "User 42", "joined": "2024-01-01", "role": "developer" }

code-review(code: "const x = eval(input)", language: "javascript")
→ "Please review this javascript for bugs, security issues, and improvements:..."

debug(error: "TypeError: Cannot read properties of undefined")
→ [user] "I'm seeing this error: TypeError: Cannot read..."
→ [assistant] "I will analyze the error and provide step-by-step debugging guidance."

Getting Started

Using as a Template

This repo is a GitHub template. Click Use this template on GitHub, then:

1. Clone your new repo

git clone https://github.com/your-username/your-repo.git
cd your-repo

2. Run the setup script

./setup.sh

setup.sh prompts for a project name and description, asks whether to keep the example code or start from a minimal skeleton, updates package.json, rewrites README.md with a clean starting point, installs dependencies, creates an initial git commit, and deletes itself.

3. Start developing

npm run dev

If you kept the examples, src/index.ts is unchanged — study the patterns and modify from there. If you started empty, you get a minimal skeleton with a single hello tool to build from.

Development Commands

Build

npm run build

Bundles src/index.ts to dist/index.js using esbuild (minified ESM output).

Dev server

npm run dev
# Test environment variables during dev:
GREETING="Howdy" npm run dev

Runs parallel watch mode: esbuild rebuilds on source changes, mctx-dev hot-reloads the server on rebuild.

Testing

npm test                          # Run all tests
npm test -- --watch               # Watch mode
npm test src/index.test.ts        # Specific file
npm test -- -t "greet"            # Pattern match

Linting and formatting

npm run lint
npm run format
npm run format:check

Environment Variables

GREETING — Customizes the greeting in the greet tool (default: "Hello"). Set GREETING="Howdy" to get "Howdy, Alice!".

Project Structure

src/index.ts        → Server implementation — all capabilities in one file
  ├─ Tools          → greet, whoami, calculate, analyze, smart-answer
  ├─ Resources      → docs://readme (static), user://{userId} (dynamic)
  ├─ Prompts        → code-review (single-message), debug (multi-message)
  └─ Export         → fetch handler for JSON-RPC 2.0 over HTTP

src/index.test.ts   → Tests for every tool, resource, and prompt
dist/index.js       → Bundled output (generated by esbuild, not committed on main)

Deployment Model

mctx does not run build commands. It serves dist/index.js from the release branch — no npm run build at deploy time.

dist/ is gitignored on main. The release pipeline (.github/workflows/release.yml) builds dist/index.js from source and commits it to release automatically. Source-only on main; built output on release.

Deployment trigger: mctx watches for version changes in package.json on the release branch. A version bump triggers a new deployment. A push with no version bump does not trigger deployment.

Conventional commits determine the version bump:

This repo uses squash merging — PR title becomes the commit subject, so PR titles must follow conventional commit format.

Do not edit the release branch directly. Do not commit dist/index.js on main.

Making Your MCP Server Discoverable

Three package.json fields and README.md determine how developers find your MCP server.

• description — Appears in the MCP Community Registry (truncates at ~100–150 chars) and on your mctx.ai page. Front-load the most important information.
• homepage — Clickable link on your public mctx.ai page. Point it at your GitHub repo or docs site.
• README.md — Becomes the documentation on your mctx.ai page and is indexed by Context7 for AI assistant discovery. Lead with what the MCP server does; the first ~4,000 characters are what AI assistants use to understand and recommend it.

Learn More

• @mctx-ai/mcp — Framework documentation and API reference
• docs.mctx.ai — Platform guides for deploying and managing your MCP servers
• mctx.ai — Host your MCP server for free
• MCP Specification — The protocol spec this MCP server implements