Latitude
Open-source observability for AI agents that doesn't just show you where an agent failed in production — it can open a pull request with Claude Code or Cursor to fix it.
🔗 Visit Latitude
Description
Watching an AI agent fail in production is only half the problem; someone still has to figure out why and fix it, often digging through logs by hand. Latitude closes that loop by combining full tracing of agent behavior with automatic failure discovery and, notably, the ability to dispatch a coding agent (Claude Code or Cursor) to draft a fix as a pull request — closer to a self-healing system than a dashboard you stare at after something breaks.
It captures conversation intelligence and behavior-pattern analysis across agent runs, automatically surfaces and prioritizes failure trends rather than requiring manual log review, and alerts via Slack, email, or webhooks when something needs attention. Tracing is OpenTelemetry-compatible, avoiding vendor lock-in, and it supports MCP and SAML SSO for enterprise deployments. It's open source (MIT license), SOC 2 Type II certified and GDPR compliant, and reports meaningful production usage: over 16 million spans ingested, 343,000+ issues detected, and 568,000+ npm downloads.
💬 Our review
The short version: against LangSmith, LangFuse, and Portkey — the other major names in LLM/agent observability — Latitude's standout feature is closing the loop from detection to fix by dispatching Claude Code or Cursor directly, which none of the others do out of the box.
Being open source and OpenTelemetry-compatible also matters: it avoids the vendor lock-in that comes with proprietary tracing formats, and the free Starter tier (20K credits, 30-day retention, unlimited seats) is generous enough for a small team to run real production monitoring before paying anything. At $99/month for 100K credits and 90-day retention, Pro is competitively priced against LangSmith and similar platforms for teams that have outgrown the free tier. The auto-fix feature is the one to sanity-check carefully — a PR opened automatically against a production failure still needs human review before merging, so it's an accelerant for debugging, not a replacement for engineering judgment. Strong pick for teams already using Claude Code or Cursor who want tracing and remediation to feel like one workflow instead of two disconnected tools.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Starter Free : 20K crédits/mois, rétention 30j, sièges illimités. Pro : 99$/mois, 100K crédits/mois, rétention 90j. Enterprise sur devis. Crédits supplémentaires : 20$/10K.
Pros
Dispatch d'agent (Claude Code/Cursor) pour corriger automatiquement les échecs via PR
Open-source (MIT) et compatible OpenTelemetry, pas de vendor lock-in
Tier gratuit généreux : 20K crédits/mois, sièges illimités
Certifié SOC 2 Type II et conforme RGPD
Usage production réel démontré : 16M+ spans, 343K+ issues détectées
Cons
Les correctifs auto-générés nécessitent toujours une revue humaine avant merge
Modèle de crédits peut devenir complexe à budgéter à grande échelle
Marché concurrentiel avec des acteurs établis (LangSmith, LangFuse)
Rétention de 30 jours sur le tier gratuit, limitée pour l'historique long terme