The open source observability layer for AI agents built on the Vercel AI SDK. Costs, latency, tokens, distributed traces, evals, and alerts for every generateText / streamText call — in two lines.

👋 I'm Gustavo, one of the makers of Foglamp. I kept shipping AI agents I couldn't actually see. Costs would creep up with no idea which agent was to blame. An agent would quietly start looping and burn tokens for an hour. Didn't find a great alternative to AI SDK, so I built one. Foglamp: observability for AI agents. Wrap your model in one line and you get, on every call: - 💸 Cost, latency & token usage — per agent, per workflow, even reasoning tokens - 🔭 Distributed traces of the whole run, with replay - ✅ Evals on your production traffic (LLM-as-judge + code checks like "No PII") - 🚨 Alerts when spend spikes or pass-rate drops It's open source (Apache 2.0) and self-hostable, there's a free hosted tier, and you can see your first trace in less than 2 minutes. Where we're headed: closing the loop from observe → act — recommending cheaper models that still pass your evals, killing runaway agents before they cost you, and optimizing prompts from your own failure traces. Would genuinely love your feedback 🙏 What's the scariest way an agent has surprised you in production?

@gustavofior Visibility is a big missing layer for agents. It’s hard to trust an autonomous workflow if you can’t see what it tried, where it failed, and what changed along the way. Agent observability will probably become a default expectation.

Foglamp

Ship AI agents you can actually see

Ship AI agents you can actually see