Fastmcp

FastMCP - Build MCP Servers in Python

FastMCP is a Python framework for building Model Context Protocol (MCP) servers that expose tools, resources, and prompts to Large Language Models like Claude. This skill provides production-tested patterns, error prevention, and deployment strategies for building robust MCP servers.

Quick Start

Installation

pip install fastmcp
# or
uv pip install fastmcp

Minimal Server

from fastmcp import FastMCP

# MUST be at module level for FastMCP Cloud
mcp = FastMCP("My Server")

@mcp.tool()
async def hello(name: str) -> str:
    """Say hello to someone."""
    return f"Hello, {name}!"

if __name__ == "__main__":
    mcp.run()

Run it:

# Local development
python server.py

# With FastMCP CLI
fastmcp dev server.py

# HTTP mode
python server.py --transport http --port 8000

What's New in v2.14.x (December 2025)

v2.14.2 (December 31, 2024)

• MCP SDK pinned to dict:

# Clients see highest version by default # Can request specific version return {"data": [...]}

### Session-Scoped State

```python
@mcp.tool()
async def set_preference(key: str, value: str, ctx: Context) -> dict:
    await ctx.set_state(key, value)  # Persists across session
    return {"saved": True}

@mcp.tool()
async def get_preference(key: str, ctx: Context) -> dict:
    value = await ctx.get_state(key, default=None)
    return {"value": value}

Other Features

• --reload flag for auto-restart during development
• Automatic threadpool dispatch for sync functions
• Tool timeouts
• OpenTelemetry tracing
• Component authorization: @tool(auth=require_scopes("admin"))

Migration Guide

Pin to v2 if not ready:

# requirements.txt
fastmcp dict:  # Use async for I/O
    async with httpx.AsyncClient() as client:
        return (await client.get(url)).json()

Resources

Expose data to LLMs. URI schemes: data://, file://, resource://, info://, api://, or custom.

@mcp.resource("user://{user_id}/profile")  # Template with parameters
async def get_user(user_id: str) -> dict:  # CRITICAL: param names must match
    return await fetch_user_from_db(user_id)

Prompts

Pre-configured prompts with parameters.

@mcp.prompt("analyze")
def analyze_prompt(topic: str) -> str:
    return f"Analyze {topic} considering: state, challenges, opportunities, recommendations."

Context Features

Inject Context parameter (with type hint!) for advanced features:

Elicitation (User Input):

from fastmcp import Context

@mcp.tool()
async def confirm_action(action: str, context: Context) -> dict:
    confirmed = await context.request_elicitation(prompt=f"Confirm {action}?", response_type=str)
    return {"status": "completed" if confirmed.lower() == "yes" else "cancelled"}

Progress Tracking:

@mcp.tool()
async def batch_import(file_path: str, context: Context) -> dict:
    data = await read_file(file_path)
    for i, item in enumerate(data):
        await context.report_progress(i + 1, len(data), f"Importing {i + 1}/{len(data)}")
        await import_item(item)
    return {"imported": len(data)}

Sampling (LLM calls from tools):

@mcp.tool()
async def enhance_text(text: str, context: Context) -> str:
    response = await context.request_sampling(
        messages=[{"role": "user", "content": f"Enhance: {text}"}],
        temperature=0.7
    )
    return response["content"]

Background Tasks (v2.14.0+)

Long-running operations that report progress without blocking clients. Uses Docket task scheduler (always enabled in v2.14.0+).

Basic Usage:

@mcp.tool(task=True)  # Enable background task mode
async def analyze_large_dataset(dataset_id: str, context: Context) -> dict:
    """Analyze large dataset with progress tracking."""
    data = await fetch_dataset(dataset_id)

    for i, chunk in enumerate(data.chunks):
        # Report progress to client
        await context.report_progress(
            current=i + 1,
            total=len(data.chunks),
            message=f"Processing chunk {i + 1}/{len(data.chunks)}"
        )
        await process_chunk(chunk)

    return {"status": "complete", "records_processed": len(data)}

Task States: pending → running → completed / failed / cancelled

When to Use:

• Operations taking >30 seconds (LLM timeout risk)
• Batch processing with per-item status updates
• Operations that may need user input mid-execution
• Long-running API calls or data processing

Known Limitation (v2.14.x):

• statusMessage from ctx.report_progress() is not forwarded to clients during background task polling (GitHub Issue #2904)
• Progress messages appear in server logs but not in client UI
• Workaround: Use official MCP SDK (mcp>=1.10.0) instead of FastMCP for now
• Status: Fix pending in PR #2906

Important: Tasks execute through Docket scheduler. Cannot execute tasks through proxies (will raise error).

Sampling with Tools (v2.14.1+)

Servers can pass tools to ctx.sample() for agentic workflows where the LLM can call tools during sampling.

Agentic Sampling:

from fastmcp import Context
from fastmcp.sampling import AnthropicSamplingHandler

# Configure sampling handler
mcp = FastMCP("Agent Server")
mcp.add_sampling_handler(AnthropicSamplingHandler(api_key=os.getenv("ANTHROPIC_API_KEY")))

@mcp.tool()
async def research_topic(topic: str, context: Context) -> dict:
    """Research a topic using agentic sampling with tools."""

    # Define tools available during sampling
    research_tools = [
        {
            "name": "search_web",
            "description": "Search the web for information",
            "inputSchema": {"type": "object", "properties": {"query": {"type": "string"}}}
        },
        {
            "name": "fetch_url",
            "description": "Fetch content from a URL",
            "inputSchema": {"type": "object", "properties": {"url": {"type": "string"}}}
        }
    ]

    # Sample with tools - LLM can call these tools during reasoning
    result = await context.sample(
        messages=[{"role": "user", "content": f"Research: {topic}"}],
        tools=research_tools,
        max_tokens=4096
    )

    return {"research": result.content, "tools_used": result.tool_calls}

Single-Step Sampling:

@mcp.tool()
async def get_single_response(prompt: str, context: Context) -> dict:
    """Get a single LLM response without tool loop."""

    # sample_step() returns SampleStep for inspection
    step = await context.sample_step(
        messages=[{"role": "user", "content": prompt}],
        temperature=0.7
    )

    return {
        "content": step.content,
        "model": step.model,
        "stop_reason": step.stop_reason
    }

Sampling Handlers:

• AnthropicSamplingHandler - For Claude models (v2.14.1+)
• OpenAISamplingHandler - For GPT models

Known Limitation: ctx.sample() works when client connects to a single server but fails with "Sampling not supported" error when multiple servers are configured in client. Tools without sampling work fine. (Community-sourced finding)

Storage Backends

Built on py-key-value-aio for OAuth tokens, response caching, persistent state.

Available Backends:

• Memory (default): Ephemeral, fast, dev-only
• Disk: Persistent, encrypted with FernetEncryptionWrapper, platform-aware (Mac/Windows default)
• Redis: Distributed, production, multi-instance
• Others: DynamoDB, MongoDB, Elasticsearch, Memcached, RocksDB, Valkey

Basic Usage:

from key_value.stores import DiskStore, RedisStore
from key_value.encryption import FernetEncryptionWrapper
from cryptography.fernet import Fernet

# Disk (persistent, single instance)
mcp = FastMCP("Server", storage=DiskStore(path="/app/data/storage"))

# Redis (distributed, production)
mcp = FastMCP("Server", storage=RedisStore(
    host=os.getenv("REDIS_HOST"), password=os.getenv("REDIS_PASSWORD")
))

# Encrypted storage (recommended)
mcp = FastMCP("Server", storage=FernetEncryptionWrapper(
    key_value=DiskStore(path="/app/data"),
    fernet=Fernet(os.getenv("STORAGE_ENCRYPTION_KEY"))
))

Platform Defaults: Mac/Windows use Disk, Linux uses Memory. Override with storage parameter.

Server Lifespans

⚠️ Breaking Change in v2.13.0: Lifespan behavior changed from per-session to per-server-instance.

Initialize/cleanup resources once per server (NOT per session) - critical for DB connections, API clients.

from contextlib import asynccontextmanager
from dataclasses import dataclass

@dataclass
class AppContext:
    db: Database
    api_client: httpx.AsyncClient

@asynccontextmanager
async def app_lifespan(server: FastMCP):
    """Runs ONCE per server instance."""
    db = await Database.connect(os.getenv("DATABASE_URL"))
    api_client = httpx.AsyncClient(base_url=os.getenv("API_BASE_URL"), timeout=30.0)

    try:
        yield AppContext(db=db, api_client=api_client)
    finally:
        await db.disconnect()
        await api_client.aclose()

mcp = FastMCP("Server", lifespan=app_lifespan)

# Access in tools
@mcp.tool()
async def query_db(sql: str, context: Context) -> list:
    app_ctx = context.fastmcp_context.lifespan_context
    return await app_ctx.db.query(sql)

ASGI Integration (FastAPI/Starlette):

mcp = FastMCP("Server", lifespan=mcp_lifespan)
app = FastAPI(lifespan=mcp.lifespan)  # ✅ MUST pass lifespan!

State Management:

context.fastmcp_context.set_state(key, value)  # Store
context.fastmcp_context.get_state(key, default=None)  # Retrieve

Middleware System

8 Built-in Types: TimingMiddleware, ResponseCachingMiddleware, LoggingMiddleware, RateLimitingMiddleware, ErrorHandlingMiddleware, ToolInjectionMiddleware, PromptToolMiddleware, ResourceToolMiddleware

Execution Order (order matters!):

Request Flow:
  → ErrorHandlingMiddleware (catches errors)
    → TimingMiddleware (starts timer)
      → LoggingMiddleware (logs request)
        → RateLimitingMiddleware (checks rate limit)
          → ResponseCachingMiddleware (checks cache)
            → Tool/Resource Handler

Basic Usage:

from fastmcp.middleware import ErrorHandlingMiddleware, TimingMiddleware, LoggingMiddleware

mcp.add_middleware(ErrorHandlingMiddleware())  # First: catch errors
mcp.add_middleware(TimingMiddleware())         # Second: time requests
mcp.add_middleware(LoggingMiddleware(level="INFO"))
mcp.add_middleware(RateLimitingMiddleware(max_requests=100, window_seconds=60))
mcp.add_middleware(ResponseCachingMiddleware(ttl_seconds=300, storage=RedisStore()))

Custom Middleware:

from fastmcp.middleware import BaseMiddleware

class AccessControlMiddleware(BaseMiddleware):
    async def on_call_tool(self, tool_name, arguments, context):
        user = context.fastmcp_context.get_state("user_id")
        if user not in self.allowed_users:
            raise PermissionError(f"User not authorized")
        return await self.next(tool_name, arguments, context)

Hook Hierarchy: on_message (all) → on_request/on_notification → on_call_tool/on_read_resource/on_get_prompt → on_list_* (list operations)

Server Composition

Two Strategies:

1. import_server() - Static snapshot: One-time copy at import, changes don't propagate, fast (no runtime delegation). Use for: Finalized component bundles.

1. mount() - Dynamic link: Live runtime link, changes immediately visible, runtime delegation (slower). Use for: Modular runtime composition.

Basic Usage:

# Import (static)
main_server.import_server(api_server)  # One-time copy

# Mount (dynamic)
main_server.mount(api_server, prefix="api")  # Tools: api.fetch_data
main_server.mount(db_server, prefix="db")    # Resources: resource://db/path

Tag Filtering:

@api_server.tool(tags=["public"])
def public_api(): pass

main_server.import_server(api_server, include_tags=["public"])  # Only public
main_server.mount(api_server, prefix="api", exclude_tags=["admin"])  # No admin

Resource Prefix Formats:

• Path (default since v2.4.0): resource://prefix/path
• Protocol (legacy): prefix+resource://path

main_server.mount(subserver, prefix="api", resource_prefix_format="path")

OAuth & Authentication

4 Authentication Patterns:

1. Token Validation (JWTVerifier): Validate external tokens
2. External Identity Providers (RemoteAuthProvider): OAuth 2.0/OIDC with DCR
3. OAuth Proxy (OAuthProxy): Bridge to providers without DCR (GitHub, Google, Azure, AWS, Discord, Facebook)
4. Full OAuth (OAuthProvider): Complete authorization server

Pattern 1: Token Validation

from fastmcp.auth import JWTVerifier

auth = JWTVerifier(issuer="https://auth.example.com", audience="my-server",
                   public_key=os.getenv("JWT_PUBLIC_KEY"))
mcp = FastMCP("Server", auth=auth)

Pattern 3: OAuth Proxy (Production)

from fastmcp.auth import OAuthProxy
from key_value.stores import RedisStore
from key_value.encryption import FernetEncryptionWrapper
from cryptography.fernet import Fernet

auth = OAuthProxy(
    jwt_signing_key=os.environ["JWT_SIGNING_KEY"],
    client_storage=FernetEncryptionWrapper(
        key_value=RedisStore(host=os.getenv("REDIS_HOST"), password=os.getenv("REDIS_PASSWORD")),
        fernet=Fernet(os.environ["STORAGE_ENCRYPTION_KEY"])
    ),
    upstream_authorization_endpoint="https://github.com/login/oauth/authorize",
    upstream_token_endpoint="https://github.com/login/oauth/access_token",
    upstream_client_id=os.getenv("GITHUB_CLIENT_ID"),
    upstream_client_secret=os.getenv("GITHUB_CLIENT_SECRET"),
    enable_consent_screen=True  # CRITICAL: Prevents confused deputy attacks
)
mcp = FastMCP("GitHub Auth", auth=auth)

OAuth Proxy Features: Token factory pattern (issues own JWTs), consent screens (prevents bypass), PKCE support, RFC 7662 token introspection

Supported Providers: GitHub, Google, Azure, AWS Cognito, Discord, Facebook, WorkOS, AuthKit, Descope, Scalekit, OCI (v2.13.1)

Supabase Provider (v2.14.2+):

from fastmcp.auth import SupabaseProvider

auth = SupabaseProvider(
    auth_route="/custom-auth",  # Custom auth route (new in v2.14.2)
    # ... other config
)

Icons, API Integration, Cloud Deployment

Icons: Add to servers, tools, resources, prompts. Use Icon(url, size), data URIs via Icon.from_file() or Image.to_data_uri() (v2.13.1).

API Integration (3 Patterns):

1. Manual: httpx.AsyncClient with base_url/headers/timeout
2. OpenAPI Auto-Gen: FastMCP.from_openapi(spec, client, route_maps) - GET→Resources/Templates, POST/PUT/DELETE→Tools
3. FastAPI Conversion: FastMCP.from_fastapi(app, httpx_client_kwargs)

Cloud Deployment Critical Requirements:

1. ❗ Module-level server named mcp, server, or app
2. PyPI dependencies only in requirements.txt
3. Public GitHub repo (or accessible)
4. Environment variables for config

# ✅ CORRECT: Module-level export
mcp = FastMCP("server")  # At module level!

# ❌ WRONG: Function-wrapped
def create_server():
    return FastMCP("server")  # Too late for cloud!

Deployment: https://fastmcp.cloud → Sign in → Create Project → Select repo → Deploy

Client Config (Claude Desktop):

{"mcpServers": {"my-server": {"url": "https://project.fastmcp.app/mcp", "transport": "http"}}}

30 Common Errors (With Solutions)

Error 1: Missing Server Object

Error: RuntimeError: No server object found at module level Cause: Server not exported at module level (FastMCP Cloud requirement) Solution: `mcp = FastM

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Kgeminic
• Source: Kgeminic/claude-skills-1
• License: MIT

Install and usage instructions live in the source repository linked above.