Launching today
AgentTrace
Open-source circuit breaker for AI agent pipelines
4 followers
Open-source circuit breaker for AI agent pipelines
4 followers
Agent 1 hallucinates. Agent 2 builds on it. Agent 3 executes. All returned 200 OK. Nobody knew. AgentTrace is the open-source circuit breaker for AI agent pipelines: - 14 rules: hallucination, PII, prompt injection, rate limiting, OWASP LLM, EU AI Act - One agent blocked = downstream agents never run - Live dashboard, zero cloud, 100% local audit trail - Python + TypeScript full parity 2 lines to add. Works with LangChain, CrewAI, OpenAI. MIT licensed. 200+ tests.
Hey Product Hunt ๐
I'm Kalash, maker of AgentTrace.
๐ง๐ต๐ฒ ๐ฝ๐ฟ๐ผ๐ฏ๐น๐ฒ๐บ: Three months ago I was debugging a multi-agent AI pipeline. Agent 1 (researcher) returned bad data, a hallucinated number. Agent 2 (drafter) built a report on it. Agent 3 (executor) sent it to customers. Every single one returned 200 OK. Nobody in my logs caught it.
We have observability for infrastructure, but zero accountability for AI agents. That's the gap AgentTrace fills.
๐ช๐ต๐ฎ๐ ๐ถ๐ ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐ฑ๐ผ๐ฒ๐:
โ Wraps any agent in 2 lines of code
โ 14 compliance rules run in parallel (PII, hallucination, prompt injection, EU AI Act...)
โ One agent violates a rule โ entire pipeline stops, downstream agents never execute
โ Full audit trail with UUIDs you can look up months later
โ Live dashboard with risk distribution, pipeline monitor, violation details
It's not just logging. It's pre-mortem intervention.
๐ง๐ต๐ฒ ๐ต๐ฎ๐น๐น๐๐ฐ๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป ๐ฑ๐ฒ๐๐ฒ๐ฐ๐๐ถ๐ผ๐ป is worth mentioning. It works without calling another LLM. Extracts numeric values, normalizes units, checks against your RAG context. Agent says "8000mg" but your source says "2000mg" โ blocked with 0.98 confidence.
๐ช๐ต๐ ๐ ๐๐ง: I want every AI team to have this, not just the ones with $50k/year security budgets. The SDK, rules engine, and dashboard are free forever.
GitHub: github.com/kalash33/agenttrace
Kalash