Launched today: Agentmetry, a local-first endpoint sensor for AI coding agents. Open source, Apache-2.0.

The problem: coding agents now run shell commands, read files and call MCP tools with the developer's own permissions. The transcript is a conversation, not a record, and EDR sees a process start a process, which is exactly what a working agent looks like.

How it works:
- Hooks at the tool boundary of Claude Code, Cursor, Codex, GitHub Copilot, Antigravity, Qwen, Kimi, Qoder and CodeBuddy. Stdio MCP servers go through an audit proxy.
- Every tool call lands in a hash-chained JSONL trail on your machine. Edit one record and agentmetry verify fails. agentmetry anchor publishes a checkpoint the host cannot rewrite.
- Arguments are hashed inside the hook by default. DLP matches record the rule id, never the secret.
- Sequence detections, not single-event alerts: a credential read followed by network egress is one finding. Labelled with MITRE ATT&CK, and ATLAS where it adds something.
- MCP rug-pull detection: tools/list is fingerprinted per tool, so a listing that moves while the config stays the same is flagged (ATLAS AML.T0109).
- Forwards to Splunk, Elastic, Google SecOps or Microsoft Sentinel. No telemetry, nothing leaves until you configure a sink.

What it does not do:
- It only sees agents whose hooks are installed. A browser chat or an IDE with hooks off is invisible.
- It records what a shell call launched, not what that process did inside. That is your EDR's job.
- Windows and Linux are tested in CI. macOS is in preview.
- Public alpha, one maintainer.

The detection benchmark ships in the package: 57 cases, 0 missed, 0 false positives. Rerun it with agentmetry benchmark rather than trusting me.

Try it with nothing installed: uvx agentmetry demo
It replays a credential theft into a temp folder, raises the finding, then shows the trail failing verification after one edit.

Criticism very welcome, especially from anyone doing detection engineering or rolling agents out to a team.

Code: https://github.com/blitzcrieg1/agentmetry
Launch: https://www.producthunt.com/products/agentmetry

#infosec #AIsecurity #opensource #DFIR #detectionengineering #MCP