Build MCP servers in Python with FastMCP. Define tools / resources / prompts, build the server, test locally, deploy to FastMCP Cloud or Docker. Use whenever the user mentions building an MCP server, exposing tools to LLMs, FastMCP, building a Claude integration, or troubleshooting FastMCP module-level server, storage, lifespan, middleware, OAuth, or deployment errors.
npx skills add https://github.com/jezweb/claude-skills --skill mcp-builder
Build a working MCP server from a description of the tools you need. Produces a deployable Python server using FastMCP.
Ask what the server needs to provide:
A brief like "MCP server for querying our customer database" is enough.
pip install fastmcp
Create the server file. The server instance MUST be at module level:
from fastmcp import FastMCP
# MUST be at module level for FastMCP Cloud
mcp = FastMCP("My Server")
@mcp.tool()
async def search_customers(query: str) -> str:
"""Search customers by name or email."""
# Implementation here
return f"Found customers matching: {query}"
@mcp.resource("customers://{customer_id}")
async def get_customer(customer_id: str) -> str:
"""Get customer details by ID."""
return f"Customer {customer_id} details"
if __name__ == "__main__":
mcp.run()
For Claude Code terminal use, add scripts alongside the MCP server:
my-mcp-server/
├── src/index.ts # MCP server (for Claude.ai)
├── scripts/
│ ├── search.ts # CLI version of search tool
│ └── _shared.ts # Shared auth/config
├── SCRIPTS.md # Documents available scripts
└── package.json
CLI scripts provide file I/O, batch processing, and richer output that MCP can't.
See assets/SCRIPTS-TEMPLATE.md and assets/script-template.ts for TypeScript templates.
Quick test -- run directly:
python server.py
Dev mode with inspector UI (recommended):
fastmcp dev server.py
# Opens inspector at http://localhost:5173
# Hot reload, detailed logging, tool/resource inspection
HTTP mode for remote clients:
python server.py --transport http --port 8000
Automated test script using FastMCP Client:
import asyncio
from fastmcp import Client
async def test_server(server_path):
async with Client(server_path) as client:
# List everything
tools = await client.list_tools()
resources = await client.list_resources()
prompts = await client.list_prompts()
print(f"Tools: {[t.name for t in tools]}")
print(f"Resources: {[r.uri for r in resources]}")
print(f"Prompts: {[p.name for p in prompts]}")
# Call first tool
if tools:
result = await client.call_tool(tools[0].name, {})
print(f"Tool result: {result}")
# Read first resource
if resources:
data = await client.read_resource(resources[0].uri)
print(f"Resource data: {data}")
asyncio.run(test_server("server.py"))
Run these checks before deploying. All required checks must pass.
Required (will cause deploy failure):
python3 -m py_compile server.py grep -q "^mcp = FastMCP\|^server = FastMCP\|^app = FastMCP" server.py
requirements.txt exists with PyPI packages only (no git+, -e, .whl, .tar.gz)api_key = "..." patterns excluding os.getenv/os.environ)Advisory (warnings):
fastmcp listed in requirements.txt.gitignore includes .env10. Server can load: timeout 5 fastmcp inspect server.py
FastMCP Cloud (simplest):
git add . && git commit -m "Ready for deployment"
git push -u origin main
# Visit https://fastmcp.cloud, connect repo, add env vars, deploy
# URL: https://your-project.fastmcp.app/mcp
Cloud requirements:
mcp, server, or apprequirements.txtDocker (self-hosted):
FROM python:3.12-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY . .
EXPOSE 8000
CMD ["python", "server.py", "--transport", "http", "--port", "8000"]
Cloudflare Workers (edge):
See the cloudflare-worker-builder skill for Workers-based MCP servers.
FastMCP Cloud requires the server instance at module level:
# CORRECT
mcp = FastMCP("My Server")
@mcp.tool()
def my_tool(): ...
# WRONG -- Cloud can't find the server
def create_server():
mcp = FastMCP("My Server")
return mcp
# FIX for factory pattern -- export at module level
def create_server() -> FastMCP:
mcp = FastMCP("server")
return mcp
mcp = create_server()
FastMCP uses type annotations to generate tool schemas:
@mcp.tool()
async def search(
query: str, # Required parameter
limit: int = 10, # Optional with default
tags: list[str] = [] # Complex types supported
) -> str:
"""Docstring becomes the tool description."""
...
Return errors as strings, don't raise exceptions:
@mcp.tool()
async def get_data(id: str) -> str:
try:
result = await fetch_data(id)
return json.dumps(result)
except NotFoundError:
return f"Error: No data found for ID {id}"
import os
from fastmcp import FastMCP
mcp = FastMCP("production-server")
API_KEY = os.getenv("API_KEY")
@mcp.tool()
async def production_tool(data: str) -> dict:
if not API_KEY:
return {"error": "API_KEY not configured"}
return {"status": "success", "data": data}
if __name__ == "__main__":
mcp.run()
These are the errors you will hit. Fix them before deploying.
| Error | Cause | Fix |
|-------|-------|-----|
| RuntimeError: No server object found at module level | Server inside a function | Export mcp = FastMCP(...) at module level |
| RuntimeError: no running event loop | Missing async/await | Use async def for async operations |
| TypeError: missing required argument 'context' | Context not type-hinted | Add context: Context with type hint |
| ValueError: Invalid resource URI | Missing URI scheme | Use data://, file://, info://, api:// |
| Resource template parameter mismatch | Name mismatch | user://{user_id} needs def get_user(user_id: str) |
| Pydantic validation error | Wrong type hints | Ensure hints match actual data types |
| Transport mismatch | Client/server protocol differ | Match both to stdio or both to http |
| Import errors with editable package | Package not installed | pip install -e . or add to PYTHONPATH |
| DeprecationWarning: mcp.settings | Old API | Use os.getenv() instead |
| Port already in use | Stale process | lsof -ti:8000 \| xargs kill -9 |
| Schema generation failure | Non-JSON types | Use JSON-compatible types (no NumPy arrays) |
| JSON serialization error | datetime/bytes in response | Convert to .isoformat() or string |
| Circular import | Factory in __init__.py | Use direct imports, avoid factory pattern |
| Python 3.12+ datetime warning | datetime.utcnow() deprecated | Use datetime.now(timezone.utc) |
| Import-time execution | Async resource at module level | Use lazy init pattern |
Keep all utilities in one file to avoid circular imports:
from fastmcp import FastMCP
import os
mcp = FastMCP("my-server")
# Config
class Config:
API_KEY = os.getenv("API_KEY", "")
BASE_URL = os.getenv("BASE_URL", "https://api.example.com")
# Helpers
def format_success(data): return {"status": "success", "data": data}
def format_error(msg): return {"status": "error", "message": msg}
@mcp.tool()
async def my_tool(query: str) -> dict:
if not Config.API_KEY:
return format_error("API_KEY not configured")
return format_success({"query": query})
Don't create async resources at module level. Initialise on first use:
_db = None
async def get_db():
global _db
if _db is None:
_db = await create_connection(Config.DB_URL)
return _db
@mcp.resource("health://status")
async def health_check() -> dict:
return {
"status": "healthy",
"version": "1.0.0",
"checks": {
"api": "connected",
"database": "connected"
}
}
import httpx
_client = None
def get_client() -> httpx.AsyncClient:
global _client
if _client is None:
_client = httpx.AsyncClient(
base_url=Config.BASE_URL,
headers={"Authorization": f"Bearer {Config.API_KEY}"},
limits=httpx.Limits(max_connections=20, max_keepalive_connections=5),
timeout=30.0
)
return _client
async def retry_with_backoff(func, max_retries=3, initial_delay=1.0):
for attempt in range(max_retries):
try:
return await func()
except Exception as e:
if attempt == max_retries - 1:
raise
delay = initial_delay * (2 ** attempt)
await asyncio.sleep(delay)
from fastmcp import Context
@mcp.tool()
async def tool_with_context(param: str, context: Context) -> dict:
# Context parameter MUST have type hint
pass
@mcp.tool()
async def long_task(items: list[str], context: Context) -> str:
for i, item in enumerate(items):
await context.report_progress(i + 1, len(items), f"Processing {item}")
await process(item)
return "Done"
@mcp.tool()
async def summarise(text: str, context: Context) -> str:
result = await context.request_sampling(
messages=[{"role": "user", "content": f"Summarise: {text}"}],
max_tokens=200
)
return result
fastmcp dev server.py # Dev mode with inspector UI
fastmcp run server.py # Run (stdio)
fastmcp run server.py --transport http --port 8000 # Run (HTTP)
fastmcp inspect server.py # Inspect without running
fastmcp install server.py # Install to Claude Desktop
fastmcp deploy server.py --name my-server # Deploy to Cloud
Environment variables: FASTMCP_LOG_LEVEL (DEBUG/INFO/WARNING/ERROR), FASTMCP_ENV (development/staging/production).
For specific integration approaches, see references/integration-patterns.md:
httpx.AsyncClient with reusable clientFastMCP.from_openapi(spec, client, route_maps=[...])FastMCP.from_fastapi(app)assets/basic-server.py -- Minimal FastMCP server templateassets/self-contained-server.py -- Server with storage and middlewareassets/tools-examples.py -- Tool patterns and type annotationsassets/resources-examples.py -- Resource URI patternsassets/prompts-examples.py -- Prompt template patternsassets/client-example.py -- MCP client usageassets/SCRIPTS-TEMPLATE.md -- CLI companion docs templateassets/script-template.ts -- TypeScript CLI script templateSkill converted from mcp-deploy-manage-agents.prompt.md
Use this skill when the user wants to launch a new AltClaw, OpenClaw, PicoClaw, or Ottie deployment through Cloud Claw. Covers the same user-facing fields and constraints exposed in the Cloud Claw UI, using the local altllm cloud-claw-* commands. Do NOT use for post-launch lifecycle tasks like start/stop/delete/logs; use cloud-claw-manage-vm.
Build hosted agents using Azure AI Projects SDK with ImageBasedHostedAgentDefinition. Use when creating container-based agents in Azure AI Foundry.
Build MCP (Model Context Protocol) servers on Cloudflare Workers with tools, resources, and prompts.
Chain agent outputs as inputs in sequential or parallel pipelines for data flow orchestration
Audit cloned or reimplemented websites for fidelity gaps, tracking scripts, source-brand and language residue, placeholders, and risky external dependencies. Use before handoff or deployment, or when asked to review a website clone for cleanup and readiness.
> Install and operate Hermes Tweet, a Hermes Agent plugin for X/Twitter research, timeline reading, tweet analysis, and approval-gated tweet actions. Use this skill when installing Hermes Tweet, researching X/Twitter accounts, monitoring launch signals, investigating mentions, auditing giveaways, or preparing guarded tweet actions. Use proactively when a Hermes Agent workflow needs current X/Twitter context. Requires XQUIK_API_KEY for read and action tools.
Handles LLM-as-judge evaluation workflows on Arize including creating/updating evaluators, running evaluations on spans or experiments, managing tasks, trigger-run operations, column mapping, and continuous monitoring. Use when the user mentions create evaluator, LLM judge, hallucination, faithfulness, correctness, relevance, run eval, score spans, score experiment, trigger-run, column mapping, continuous monitoring, or improve evaluator prompt.
Take jezweb/mcp-builder from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference pip.
Without those the skill loads but fails at the first command.