my-mcp-template

my-mcp-template

A generic template for building FastMCP serversfor CLI coding agents (Claude Code and similar) -- not chat UIs. It shipswith two working example tools (pure logic, and an external-call tool withtenacity retries), a unified error format, SSO/OIDC auth, optional Langfusetool-call tracing, Docker Compose deployment, and a two-axis test suite.Not RAG-specific: just as suited to DB-query or file-operation tools.

Using this as a template

This repo is a plain [GitHub template repository]("Use this template"button), not a cookiecutter project -- there's no templating variables tofill in. To start a new project from it:

1. Click "Use this template" on GitHub (or git clone + re-init).
2. Rename the package path if you want something other thansrc.main.python -- it's a literal directory structure(src/main/python/), so a plain find-replace across the repo handles it.
3. Replace example_tool.py with your own tools (seedocs/TOOL_GUIDELINES.md).

Local development

uv sync
cp .env.example .env          # AUTH_ENABLED=false by default, fine for local dev
uv run python -m src.main.python.main

The server listens on http://0.0.0.0:8000/mcp (streamable HTTP,stateless). Point MCP Inspectoror any streamable-HTTP-capable client at that URL to try it out.

Running with Docker Compose

cp .env.example .env
docker compose up --build

Health check: curl http://localhost:8000/health.

Tests

Two independent groups, plus a fast/slow cost axis that cuts across both:

• tests/server/ -- does the MCP server itself work correctly?Schema/description contract, golden-case functional correctness, authverification logic, and a real container integration smoke test.
• tests/agent/ -- can an agent actually use it? Feeds your tooldescriptions to a real Claude model and checks it picks the right tool.This is the layer that actually tests description quality.

uv run pytest -m "not slow"   # fast: no LLM calls, no docker. Run on every PR.
uv run pytest -m slow          # slow: calls Anthropic + docker compose. Run on merge to main.
uv run pytest                  # everything

ANTHROPIC_API_KEY is required for tests/agent/ (fixed to Claude onpurpose -- see below). LLM_JUDGE_BASE_URL/LLM_JUDGE_API_KEY/LLM_JUDGE_MODELare required for llm_judge golden cases -- bring your own OpenAI-compatibleprovider (OpenAI, NVIDIA NIM, DeepSeek, Together, a local vLLM/Ollama, ...).Without them, llm_judge cases skip with a clear reason rather than fail.

Golden cases

Add your own in tests/golden/*.yaml. Six assert_types:exact_match, contains, regex_match, numeric_tolerance, llm_judge,custom -- see tests/golden/schema.py for the shape andtests/golden/example.yaml for one of each.

Faithfulness layer (optional, deepeval)

For RAG-style tools where you want to check answers stay grounded inretrieved context, add deepevalassertions in your own test module (the eval dependency group is alreadyset up: uv sync --group eval). Not wired into a specific test file heresince it only applies if your tools actually do retrieval.

Connecting a client

{
  "mcpServers": {
    "my-mcp-template": {
      "url": "http://localhost:8000/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

The Authorization header is only required when AUTH_ENABLED=true. Seedocs/DEPLOYMENT.md for how tokens get issued in production (SSO/OIDC).

Project structure

src/main/python/    server code (main.py, config.py, auth.py, errors.py, tools/)
tests/server/        does the server work correctly?
tests/agent/          can an agent actually use it?
tests/golden/         golden case schema + data
docs/                  TOOL_GUIDELINES.md, DEPLOYMENT.md

Checklist: what you still need to fill in

Blocking (fail-fast if missing):

• Real golden case content in tests/golden/*.yaml for your own tools.
• Production OIDC_ISSUER / JWKS_URL / AUDIENCE / TENANT_CLAIM_NAME-- there's no universal default, every IdP is different.

Conditional (only if you add the matching feature):

• A volumes: entry in docker-compose.yml, if you add a realfile-operation tool.
• LLM_JUDGE_BASE_URL / LLM_JUDGE_API_KEY / LLM_JUDGE_MODEL, only ifyou use the llm_judge assert_type in a golden case -- no universaldefault, bring your own OpenAI-compatible provider. Cases skip (not fail)if unset.
• LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY / LANGFUSE_BASE_URL, onlyif you want tool-call tracing. Unset = no middleware added.

Already have a reasonable default -- tune if needed:

• LLM_JUDGE_THRESHOLD in config.py.
• ALLOWED_ORIGINS (empty = CORS off; only needed for browser clients).
• Whether to upgrade past resource-server-only auth to OAuthProxy(see docs/DEPLOYMENT.md) -- not needed unless your MCP client has noother way to obtain a token.