FastMCP - FastAPI MCP Framework & Runner

fast-mcp

FastAPI-native Model Context Protocol (MCP) framework with automatic route reflection, ASGI scope bridging, dynamic progressive tool discovery, resilient error recovery, and interactive in-chat MCP Apps (SEP-1865).

---

Highlights

• ⚡ Hybrid Dual-Citizen ASGI Mount: Mounts directly onto any existing FastAPI instance in-process over standard ASGI—zero external proxying, zero hanging subprocesses.
• 🔌 Local stdio CLI Runner & Desktop AI Bridge: Run fast-mcp stdio main:app or python -m fast_mcp stdio main:app to connect desktop AI clients (Claude Desktop, Cursor) over standard I/O pipes with zero network setup.
• 🔍 Route Reflection: Opt-in tags (tags=["mcp"]) automatically convert FastAPI endpoints, Pydantic models, docstrings, path/query/body parameters into MCP tools.
• 🛠️ Custom AI Tools (@mcp.tool): Define AI-tailored composite tools alongside reflected routes with automatic schema and docstring extraction.
• 🔐 ASGI Scope Bridging: Client authorization headers (Authorization: Bearer <token>, cookies, API keys) captured during the MCP handshake are bridged into an in-memory ASGI Request, natively resolving FastAPI's Depends() and Security() providers without code changes.
• 🧠 Dynamic Progressive Tool Discovery: Protect agent context windows via progressive discovery (dynamic_discovery=True), the search_tools(query: str) meta-tool, and zero-dependency KeywordTagRouter.
• 🛡️ Resilient Error Recovery & Minified JSON: Traps route HTTPException and Pydantic validation errors into informative CallToolResult(isError=True) responses so LLMs can self-correct without protocol failures. Output defaults to compact, token-conscious minified JSON with custom @mcp.serializer formatting hooks.
• 🖥️ Dual UI & In-Chat MCP Apps (SEP-1865): Embedded browser inspector at /mcp/docs plus native support for in-chat interactive iframes in desktop AI clients (Claude Desktop, Cursor, VS Code) via _meta.ui.resourceUri, the built-in inspect() tool, and the @mcp.app() decorator.

---

Installation

pip install mcp-fastapi

Or using uv:

uv add mcp-fastapi

---

Quickstart

from fastapi import FastAPI, Depends, Header, HTTPException
from pydantic import BaseModel, Field
from fast_mcp import FastMCP

app = FastAPI(title="Store API")
mcp = FastMCP(app=app, name="store-mcp")

# 1. Existing FastAPI route reflected automatically via tags=["mcp"]
class Product(BaseModel):
    id: int
    name: str
    price: float

@app.get("/products/{product_id}", tags=["mcp"])
async def get_product(product_id: int) -> Product:
    """Fetch product details by ID."""
    if product_id == 404:
        raise HTTPException(status_code=404, detail="Product not found")
    return Product(id=product_id, name="Smart Widget", price=29.99)

# 2. Custom AI tool with native dependency injection
def verify_token(authorization: str = Header(...)) -> str:
    if not authorization.startswith("Bearer "):
        raise HTTPException(status_code=401, detail="Invalid token")
    return authorization.split(" ")[1]

@mcp.tool(name="order_status", description="Check customer order status")
def check_order(order_id: str, user: str = Depends(verify_token)) -> dict:
    return {"order_id": order_id, "customer": user, "status": "Shipped"}

# 3. Mount MCP endpoints (/mcp/sse, /mcp/messages, /mcp/docs)
mcp.mount()

Run with standard ASGI servers:

uvicorn main:app --reload

---

Core Capabilities

1. Route Reflection

Routes tagged with tags=["mcp"] (configurable via route_tag) are automatically inspected upon mcp.mount():

• Endpoint docstrings (Google, Sphinx, NumPy format) become tool descriptions and parameter docs.
• Pydantic request models, query parameters, and path variables become MCP input schemas.
• Untagged endpoints remain standard HTTP routes and are never leaked to LLMs.

@app.post("/items/create", tags=["mcp"])
async def create_item(item: ItemModel) -> ItemModel:
    """Create a new catalog item.

    Args:
        item: The catalog item specification.
    """
    return item

2. Custom AI Tools (@mcp.tool)

Register AI-specialized tools that don't need dedicated REST endpoints:

# Bare decorator
@mcp.tool
def calculate_quote(quantity: int, discount: float = 0.0) -> float:
    return quantity * 100.0 * (1.0 - discount)

# Parameterized decorator
@mcp.tool(name="inventory_lookup", description="Lookup stock levels", tags=["inventory"])
async def check_inventory(sku: str) -> dict:
    return {"sku": sku, "in_stock": True, "count": 42}

3. ASGI Scope Bridging & Native Auth

Incoming headers (Authorization: Bearer ..., cookies, API keys) from the MCP client's SSE handshake or message posts are captured into an active request context:

from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials

security = HTTPBearer()

@mcp.tool
async def user_profile(creds: HTTPAuthorizationCredentials = Depends(security)) -> dict:
    token = creds.credentials
    return {"user": "alice", "token_verified": True}

If authorization fails or headers are omitted, fast-mcp unwraps the resulting HTTPException(401) into CallToolResult(is_error=True) so the agent receives an actionable authentication error rather than crashing the transport.

4. Dynamic Progressive Tool Discovery

Prevent LLM context window bloat on large FastAPI applications with hundreds of endpoints:

mcp = FastMCP(
    app=app,
    dynamic_discovery=True,  # Or set dynamic_discovery_threshold=20
    baseline_tools=["search_tools", "get_system_status"],
    baseline_tag="baseline",
)

• When active, tools/list exposes only baseline tools plus the search_tools(query: str) meta-tool.
• Calling search_tools(query="invoice") executes the pluggable ToolRouter (defaults to zero-dependency KeywordTagRouter with tokenized name/tag/description ranking) and returns matching tool definitions with full JSON schemas.

5. Resilient Error Interception & Custom Serializers

• Exception Traps: HTTPException (400, 404, 422) and Pydantic validation errors return clean, concise messages with isError=True.
• Minified Output: Responses serialize to compact minified JSON ({"id":1,"name":"widget"}) saving prompt tokens.
• Custom Serializers: Format return types into tailored markdown or summaries:

class Report(BaseModel):
    title: str
    metrics: dict[str, int]

@mcp.serializer(Report)
def format_report(report: Report) -> str:
    md = f"### {report.title}\n"
    for k, v in report.metrics.items():
        md += f"- **{k}**: {v}\n"
    return md

6. Dual UI: Browser Inspector & In-Chat MCP Apps (SEP-1865)

Embedded Browser Inspector

Open http://localhost:8000/mcp/docs in any browser to inspect registered tools, view schemas, and execute test invocations interactively without external Node.js CLIs. (Disable with FastMCP(app, enable_ui=False)).

In-Chat MCP Apps (SEP-1865)

Render rich interactive HTML/JS widgets directly in modern desktop AI clients (Claude Desktop, Cursor, VS Code):

# Built-in server inspector tool
# Agent calling `inspect()` receives an interactive iframe pointed to ui://fast-mcp/inspector

# Authoring custom in-chat widgets:
@mcp.app(
    name="dashboard",
    resource_uri="ui://store/dashboard",
    html="""
    <div style="font-family: sans-serif; padding: 1rem; border-radius: 8px; background: #f0f4f8;">
        <h2>Store Live Metrics</h2>
        <p>Active Users: <strong>1,420</strong></p>
    </div>
    """
)
def live_dashboard() -> str:
    return '<iframe src="ui://store/dashboard" width="100%" height="400"></iframe>'

7. Local stdio CLI Runner & Desktop AI Bridge

Connect desktop AI clients (Claude Desktop, Cursor) directly to your FastAPI backend or FastMCP instance over standard input/output (stdio) pipes with zero network setup, port conflicts, or external proxying.

Command-Line Usage

# Run stdio runner pointing to FastAPI app or FastMCP instance
fast-mcp stdio main:app

# Or via Python module invocation
python -m fast_mcp stdio main:app

# Target attribute defaults to 'app' or 'mcp' if omitted:
fast-mcp stdio main

# The 'stdio' subcommand can also be omitted as default:
fast-mcp main:app

Claude Desktop Configuration (claude_desktop_config.json)

Configure Claude Desktop to launch your FastMCP server directly:

{
  "mcpServers": {
    "my-fast-mcp-app": {
      "command": "fast-mcp",
      "args": ["stdio", "main:app"]
    }
  }
}

Or using uv to manage the virtual environment automatically:

{
  "mcpServers": {
    "my-fast-mcp-app": {
      "command": "uv",
      "args": ["run", "fast-mcp", "stdio", "main:app"]
    }
  }
}

Programmatic Stdio Runner

You can also run stdio mode programmatically from Python:

import asyncio
from fast_mcp import FastMCP, run_stdio

mcp = FastMCP(name="my-stdio-server")

@mcp.tool()
def add(a: int, b: int) -> int:
    return a + b

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

---

Testing & Verification

fast-mcp exercises external behavior across the ASGI Protocol Seam using httpx.AsyncClient with ASGITransport:

# Run full test suite
pytest

# Run tests with coverage
pytest --cov=fast_mcp --cov-report=term-missing

---

Specification & Architectural Documents

• Interactive Documentation Website
• AI Agent Index (llms.txt)
• Specification: fast-mcp Core Framework (V1)
• GLOSSARY.md
• ADR 0001: Architecture Foundation and Hybrid Scope
• ADR 0002: Dual-UI, ASGI Scope Bridging, and Resilient Error Handling

---

License

MIT