Tracea
A free, self-hosted dashboard that shows you exactly what your AI agents did, why they cost what they cost, and where they broke.
🔗 Visit Tracea
Description
If you've ever built something with an AI assistant that acts on its own — booking things, writing code, answering customers — you've probably had the moment where it does something weird and you have no idea why. Tracea is like a flight recorder for that: it quietly logs every step the AI takes so you can rewind and see exactly what happened.
Tracea is an open-source, self-hosted observability tool for AI agents. It intercepts LLM calls and tool executions at the transport level (no SDK wrappers required) and renders them in a local React dashboard showing cost trends, token usage, duration, session health, and per-agent breakdowns. YAML-configurable alert rules catch tool errors, cost spikes, rate limits, infinite loops and empty responses, routing them to Slack or any webhook with deduplication. Root-cause analysis can run through OpenAI, Anthropic, or fully offline via Ollama, so nothing has to leave your infrastructure.
💬 Our review
The short version: Tracea is a solid pick if you're already running agents in production and want to see inside the black box without sending your traces to yet another SaaS vendor.
Its real differentiator against cloud competitors like LangSmith or Helicone is the deployment model — everything runs on your infrastructure, with no per-event pricing and no proprietary dashboard you can't inspect. That matters if you're handling sensitive data or just tired of observability bills scaling with usage. The tradeoff is that you're on the hook for hosting and updates yourself, and with only 7 GitHub stars and no listed launch date, this is clearly an early, small project — expect rough edges, smaller community support, and slower feature velocity than an established player like PandaProbe or a funded SaaS. For a solo dev or small team comfortable self-hosting, it costs nothing but time; for a team that wants guaranteed support, that same self-hosted nature is the risk.
💰 Pricing
📊 Global score
🤖 AI-enriched data
MIT license, self-hosted, $0 — no paid tier, no per-event billing
Pros
Fully self-hosted, zero vendor lock-in, no per-event pricing
Transport-level tracing — no SDK wrappers to install
Local root-cause analysis via Ollama, fully offline-capable
YAML-configurable alerting with Slack/webhook routing
Cons
Very early project (7 GitHub stars), no track record yet
You host and maintain it yourself — no managed cloud option
No listed launch date or company behind it
Smaller community than funded competitors like PandaProbe