fastmcp-langfuse

Plug Langfuse observability into any FastMCP v3 server. One function call, zero boilerplate.

from langfuse import get_client
get_client()

from fastmcp import FastMCP
from fastmcp_langfuse import instrument

mcp = FastMCP("My Server")
instrument(mcp)

@mcp.tool()
def anything(x: str) -> str:   # fully traced, just like that
    return "hello"

Three imports, one instrument() call. Vibecode the rest -- every tool, resource, and prompt gets session tracking, I/O capture, environment tags, and trace URLs automatically.

Install

pip install fastmcp-langfuse
# or
uv add fastmcp-langfuse

You'll also need langfuse in your project -- this library sets the right OTel span attributes that the Langfuse SDK picks up, but doesn't import it directly.

What you get for free

Once you call instrument(mcp), every handler registered after it gets:

• Input/output -- function args serialised as input, return value as output
• Session & user tracking -- session_id and client_id from FastMCP context
• Environment & release tags -- tag your traces with production, staging, etc.
• Trace URLs -- clickable Langfuse links printed to your console

No decorators on every handler. No manual span enrichment. Just works.

Configure it

Pass kwargs directly or use a config object -- whatever feels right:

# kwargs
instrument(mcp, environment="production", release="v1.2.3", trace_urls=False)

# or config object
from fastmcp_langfuse import LangfuseConfig

instrument(mcp, config=LangfuseConfig(environment="production"))

Need more control?

Use @langfuse_trace on individual handlers instead of (or alongside) instrument():

from fastmcp_langfuse import langfuse_trace

@mcp.tool()
@langfuse_trace
def special_tool(x: str) -> str:
    return "traced with the decorator"

Same enrichment, just explicit.

How it works under the hood

FastMCP auto-creates OTel spans for every tool/resource/prompt call. This library wraps your handlers to enrich those spans with Langfuse-specific attributes (langfuse.session.id, langfuse.trace.input, etc.) that the Langfuse SDK recognises when exporting.

The library only depends on opentelemetry-api and fastmcp. Lightweight, no opinions about your config management.

Examples

Check out examples/ for a full working demo with tools, resources, prompts, cost tracking, user feedback, and distributed tracing. It uses pydantic-settings to bridge env vars to LangfuseConfig -- but that's just one way to do it, bring your own config however you want.

cd examples && cp .env.example .env
# add your LANGFUSE_PUBLIC_KEY and LANGFUSE_SECRET_KEY

uv run python examples/server.py   # terminal 1
uv run python examples/client.py   # terminal 2

License

MIT