DEV Community

Jonathan Martin Paez
Jonathan Martin Paez

Posted on • Originally published at github.com

Lookspan: local-first observability for AI agents

Most LLM observability tools are SaaS — your prompts leave your machine and you pay per event. Lookspan is the opposite: one command, runs locally, your data never leaves your box, infra cost zero.

npx lookspan   # → http://127.0.0.1:3100
Enter fullscreen mode Exit fullscreen mode

It ingests spans/traces from your agents into a local SQLite database and shows them in a real-time dashboard:

  • a timeline (waterfall) of where time goes, plus a conversation transcript of each prompt/response
  • cost tracking per span and trace, latency p50/p95/p99
  • alerts on errors / cost / latency thresholds

It is MCP-native, with drop-in wrappers for the OpenAI and Anthropic SDKs (observeOpenAI / observeAnthropic) and an OpenTelemetry receiver — point any OTel exporter at it, no Lookspan SDK required.

Newer additions:

  • Replay a captured prompt against another model and diff cost / latency / output
  • LLM-as-judge scoring of a trace
  • Datasets to run a whole test set in batch and compare runs (model A vs B)

Local-first by design: binds to 127.0.0.1, redacts secret-looking values server-side, and your prompts/outputs never leave your machine.

MIT, TypeScript.

npx lookspan
Enter fullscreen mode Exit fullscreen mode

Repo: https://github.com/JoniMartin27/lookspan

Top comments (0)