Langfuse already holds the answer to why a run cost what it cost; the friction is going to look. Pulling the trace and the metrics into the session where the agent is being fixed puts the evidence next to the code.
A monitoring server over Langfuse. It watches instrumented agents, reads a single trace in detail for a failed execution, returns latency, cost and token metrics, and analyses performance across a time period — including comparing agent versions through metadata filters.
- Instrumented agents monitored in real time
- A detailed trace read for one execution, for debugging a failure
- Latency, cost and token usage metrics
- Performance analysed across a time period
- Agent versions compared through metadata filters
- Stdio for Claude Desktop and Cursor, or Streamable HTTP and SSE for anything else
Python 3.11 or higher, a Langfuse account, and agents already instrumented to send traces there — this reads Langfuse, it does not instrument anything for you. LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY and LANGFUSE_HOST go in the environment or a .env file. For a URL-based connection, start it in Streamable HTTP mode and point the client at its local /mcp endpoint.