Hi Product Hunt, I'm Ioannis, and I build Agentmetry on my own, from Greece.

It started with a simple question I could not answer: what did my coding agent actually do on my laptop today? The transcript is a conversation, not a record. My editor keeps no log. Endpoint security sees a process start a process, which is what an agent looks like when it is working fine.

So Agentmetry sits at the tool boundary. Claude Code, Cursor, Codex, Copilot and five more agents call it through their own hooks, and every tool call lands in a hash-chained trail on your machine. Fourteen sequence rules look for risky orders of events, like a credential read followed by a network upload, and label them with MITRE ATT&CK. If you run Splunk, Elastic, Google SecOps or Sentinel, it forwards there. Arguments are hashed inside the hook by default, so the trail holds fingerprints, not your secrets.

What it does not do, so you know before you install it:

- It only sees agents whose hooks are installed. A browser chat or an IDE with hooks off is invisible to it.

- Windows and Linux are tested in CI. macOS runs the test suite but is still in preview.

- It is public alpha, and Sentinel has not been run against a live workspace yet.

The detection benchmark is public: 57 cases, 0 missed, 0 false positives, and you can rerun it with agentmetry benchmark.

To try it without installing anything: uvx agentmetry demo. It replays a credential theft into a temp folder, catches it, then shows the trail failing verification after one edit.

I would love your hardest criticism, especially from anyone who runs a SOC or ships agents to a team.