Production-Ready FastMCP Server

MCP Server (FastMCP) and Client

Production-ready FastMCP server and a production-grade client with structured logging, env config, health checks, metrics, and containerization.

Features

• Server: Stdio and SSE runtimes via FastMCP, CORS & security headers, token auth, basic rate limiting

• Client: SSE and stdio transports, CLI to list tools/call tools/get resources, structured logs

• Structured JSON logging with structlog

• Env-based configuration

• Health endpoints and CLI checks

• Prometheus metrics primitives

• Dockerfile and Makefile

Requirements

• Python 3.9+

Setup

ls -la .venv
python3 -m venv .venv
.venv\Scripts\activate
 python -m pip install --upgrade pip
# . .venv/bin/activate   for linux
# pip install -U pip
pip install -e .[dev]

Copy and adjust environment:

cp .env.example .env || true

Run (stdio)

mcp-server-stdio

Run (SSE)

mcp-server-sse  # uses HOST, PORT, AUTH_TOKEN, CORS_ORIGINS

Health

mcp-server-health

Docker

docker build -t mcp-server:latest .
docker run --rm -p 8000:8000 -e AUTH_TOKEN=changeme mcp-server:latest

Client CLI

Environment (SSE example):

export MCP_CLIENT_TRANSPORT=sse
export MCP_SSE_URL=http://localhost:8000/sse
export AUTH_TOKEN=changeme  # if server requires it

List tools:

mcpx list-tools

Call tool:

mcpx call-tool add --args '{"a": 1, "b": 2}'

Get resource:

mcpx get-resource time://now

Health check:

mcpx health

Security

• Set a strong AUTH_TOKEN in production for SSE mode

• Restrict CORS_ORIGINS to trusted origins

• Run the container as non-root (Dockerfile does)

• Prefer TLS for SSE (VERIFY_TLS=1)

• Limit client network egress in production and rotate tokens regularly

RAG (Postgres + pgvector)

• Set DATABASE_URL (or PG* envs) and OPENAI_API_KEY.

• Enable vector extension in Postgres (the app will attempt to create it).

Ingest files via CLI:

python -m rag.cli ingest path/to/dir path/to/file.pdf

Ask a question via CLI:

python -m rag.cli ask "What does the document say about refunds?"

Query with citations via client:

mcpx rag-query "What does the document say about refunds?" --server http://localhost:8000

HTTP endpoints (when server running):

• POST /rag/upload (multipart form with files)

• POST /rag/query JSON { "question": "..." }

• GET /rag/chunk/{chunk_id} (get chunk metadata)

MCP tool:

• rag_ask(question: str) -> str

OpenTelemetry Tracing

Enable distributed tracing with:

export OTEL_ENABLE=1
export OTEL_SERVICE_NAME=mcp-server
export OTEL_EXPORTER_OTLP_ENDPOINT=http://jaeger:14268/api/traces

Traces include:

• RAG ingestion: file parsing, chunking, embedding, DB operations

• RAG retrieval: vector search, BM25 reranking, context assembly

• LLM calls: token usage, model info, latency

• Database operations: SQL queries, connection pooling