Spanlight documentation
Trace every LLM call, find what is wrong, test changes before they ship and run it all on your own Postgres. Pick a path below, or search.
- About 10 minutesSelf-host and send a traceStart the stack with Docker Compose, add three lines to a Python app and watch the trace arrive.Quickstart →
- Nothing to installTry the live demoOpen the Spanlight home page and choose Try the live demo to explore a seeded workspace. Self-hosted servers show it only when the demo is turned on.Go to the home page →
- Any languageSend traces with OpenTelemetryPoint an OTLP exporter at /v1/otlp/traces. JSON or protobuf; gen_ai.* attributes are mapped.OpenTelemetry →
Explore the product
Trace and debug
Every call with its prompt, completion, latency, tokens and cost, grouped into traces and sessions.
The Doctor
Detectors read your traces and open an insight with evidence, a fix and a way to check it.
Evaluate
Version prompts, build datasets, score answers and fail the pull request when quality drops.
Gateway
One base URL for OpenAI and Anthropic, with retries, caching and budgets that block.
Alerts and budgets
Get told when errors, latency or spend cross a line, and stop spend at a limit.
Self-hosting
One image, three roles, Postgres as the only datastore.
API and security
Every HTTP endpoint, and how Spanlight protects your data and takes vulnerability reports.