Question 1

What is Foglamp?

Accepted Answer

Observability for AI agents built on the Vercel AI SDK. Every generateText and streamText call becomes a trace with its cost, latency, tokens, prompt, and response. Evals score what your agents say, and alerts watch cost, errors, and pass rates. It is open source and you can run it yourself.

Question 2

Which AI SDK versions are supported?

Accepted Answer

Versions 4 through 6 use wrap() from foglamp/wrap around your model calls. Version 7 uses the native telemetry integration: registerTelemetry(foglamp()) once at startup. The setup prompt reads your installed version and picks the right path.

Question 3

Can I self-host it?

Accepted Answer

Yes. Foglamp is Apache 2.0 and ships with a docker compose file that brings up the app, Postgres, and ClickHouse. Nothing in the self-hosted version phones home.

Question 4

Where are my prompts and responses stored?

Accepted Answer

In ClickHouse, scoped to your project, for as long as your plan retains data: 3 days on Free and 14 days on Pro. If you self-host, all of it stays on your own infrastructure. PII redaction in the SDK is on the roadmap.

Question 5

What does the SDK add to my request latency?

Accepted Answer

Nothing on the request path. Spans are batched in memory and flushed in the background every five seconds. On serverless, call flush() before the function returns. If Foglamp is unreachable, your agent keeps running and the batch is dropped.

Know what your agents are doing

Know where your money goes

See what happened

Catch bad answers

Questions

Your first trace
is one prompt away.