foglamp, and traces stream to a backend you can run yourself or use
as a hosted service.
Foglamp meets the AI SDK where it is. On the stable v4, v5, and v6 lines
it wraps the module with
foglamp/wrap; on v7 (beta) it uses
the native telemetry-integrations API. Same traces, same backend — the
Quickstart helps you pick your path.Why Foglamp
Generative apps fail in ways traditional APM never had to model: a single user action fans out into multiple model calls, tool executions, and retries, each with its own cost and latency. Foglamp is built around that shape.Cost at ingest
Every span is priced as it arrives, broken down by input, output, reasoning,
and cached tokens. Unknown models surface as
—, never a misleading $0.Distributed traces
One top-level
generateText/streamText call is a trace; its steps and tool
calls are spans. Group them into agents, workflows, and runs.Latency & TTFT
p50/p95/p99 latency and time-to-first-token, read straight from the SDK’s
own performance metrics rather than guessed.
Self-host or hosted
Run the whole stack with
docker compose up, or point the SDK at a managed
endpoint. Same code either way.How it fits together
- SDK — instruments your AI SDK calls, batching spans and flushing them fire-and-forget. Silent no-op when no API key is set; never throws, never adds latency to your model calls.
- Ingest API — authenticates the API key, prices each span, and writes to ClickHouse.
- Dashboard — traces, workflows, agents, sessions, cost-over-time, evals, and alerts. See the dashboard tour.
Next steps
Quickstart
Install the SDK and see your first trace in minutes.
Data model
Understand traces, spans, workflows, runs, and sessions.
SDK reference
Every configuration option and integration method.
Dashboard tour
Traces, workflows, agents, sessions, evals, cost, and alerts.
Self-hosting
Run the full stack on your own infrastructure.
Troubleshooting
Not seeing traces? The usual causes and fixes.