Show HN: AgTrace – Observability for AI Coding Agents via MCP (Claude Code etc.)

2 points

5 months ago

AI agents are getting more capable, but we're increasingly in the dark about what they're actually doing. They run complex multi-step workflows, call dozens of tools, reason through problems - and we just watch the output scroll by. It's a black box, and humans end up being led around by the agent rather than understanding it.

I wanted to flip this. The key insight: all these agents (Claude Code, Codex, Gemini) already write detailed logs. The problem is they're in different locations, different formats, incompatible schemas.

agtrace normalizes this "observation layer" across providers:

- Auto-discovers logs from Claude, Codex, Gemini - Converts them into a unified event timeline - Exposes this via CLI, TUI dashboard, and MCP

The MCP part is what makes it interesting for agents themselves. An agent can now query its own past sessions:

- "What approach did we take when we refactored auth last week?" - "Show me errors from yesterday's session" - "How did we handle this edge case before?"

This enables agent self-reflection - using execution history to inform current decisions.

Built in Rust for safety and speed. 100% local, no cloud dependencies. The database is just a pointer index to original logs - rebuilable anytime.

Happy to discuss the architecture or use cases.