Show GN: Spanlens - An Open-Source Observability Platform for LLM Calls and Agent Traces in One Place

Spanlens, an open-source observability platform that integrates LLM call logging, cost tracking, and agent tracing, has been released. It focuses on solving the difficulty of tracking costs incurred from LLM usage and the complexity of debugging agents that involve multiple LLM calls.

Key Features

• Easy proxy integration: Simply change the baseURL of the OpenAI, Anthropic, or Gemini SDK to the Spanlens proxy to automatically log requests, responses, token usage, and costs. Streaming responses and Tool calling are supported identically to the original.
• Agent trace and LangGraph visualization: Traces are visualized not only as a chronological timeline but also on actual graph nodes, allowing you to see per-node latency and edge flow of LangGraph-based agents at a glance.
• Automatic Critical Path analysis: Automatically identifies the call chain that causes the most latency across the entire trace, improving debugging efficiency.
• Statistical prompt A/B testing: Compares performance differences between prompt versions based on statistical significance using a Welch t-test rather than simple averages.
• Self-hosting support: Provided under the MIT license, and can be deployed to your own server via Docker.

Tech Stack Built on Next.js 14, Hono, Supabase, and ClickHouse, using ClickHouse asynchronously for large-scale log processing to prevent data loss.