Foglamp
Drop-in observability for AI agents — tracks the cost, latency, and output quality of every LLM call so a degrading agent gets caught before users notice.
🔗 Visit Foglamp
Description
The way a car's dashboard warns you something's wrong before the engine actually fails, Foglamp is meant to sit behind an AI agent and flag problems before they reach a user — without anyone having to manually read through logs. If you've built something on top of Claude or another LLM and have no real visibility into what it's actually doing, costing, or getting wrong once it's live, that's the gap this fills.
Concretely, Foglamp monitors LLM calls to track cost, latency, and output quality over time, and is built specifically to plug into the Anthropic Claude Agent SDK and Claude Code — setup is a single prompt pasted into Claude Code that wires up the instrumentation automatically, rather than a manual SDK integration. It provides a dashboard for inspecting individual calls, configurable alerts, and evaluation runs (evals) to catch quality regressions. All plans include unlimited agents, workflows, and traces; what's actually rationed is data volume (spans) and how long that data is retained.
💬 Our review
The short version: Foglamp's one-prompt setup for Claude Code is a genuinely convenient onboarding trick, and unlimited agents/workflows/traces even on the free tier is generous — but a 3-day data retention window on that free tier is short enough to matter, and there's no self-hosted option for teams wary of another vendor holding their traces.
Against LangSmith or Helicone, which are LLM-provider-agnostic, Foglamp's tight integration with the Claude Agent SDK is both its strength and its limitation: onboarding is faster if you're already Claude-based, but it's a narrower bet if your stack spans multiple model providers. Against Langfuse, which is open source and self-hostable, Foglamp is closed and cloud-only — a real consideration for teams that want to keep trace data in-house. The $49/month Pro tier is straightforward and includes a notable jump in span volume (10,000 to 1,000,000/month) and retention (3 to 14 days), which is reasonable if the free tier's short retention becomes a real constraint. Best fit: teams building specifically on Claude/Anthropic tooling who want fast setup and don't need self-hosting. Weaker fit: multi-provider LLM shops, or anyone requiring on-premise data retention.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Free : 10 000 spans/mois, rétention 3 jours, 1 projet, 1 alerte, 5 evals, agents/workflows/traces illimités. Pro : 49 $/mois — 1 000 000 spans/mois, rétention 14 jours, 5 projets, 10 alertes, 20 evals, assistant Foggy AI, alertes email/Slack, support prioritaire. Enterprise : sur devis — rétention 90+ jours, SSO/SAML, audit logs, support dédié.
Pros
Setup en un seul prompt copié dans Claude Code
Agents, workflows et traces illimités même au palier gratuit
Prix Pro clair et fixe (49 $/mois)
Alertes et evals inclus dès le palier gratuit
Cons
Rétention gratuite très courte (3 jours)
Positionnement très centré sur l'écosystème Claude/Anthropic
Pas d'option self-hosted, contrairement à Langfuse