Agentmetry

Agentmetry is an open source security monitoring tool built specifically for AI coding agents. It captures every tool execution from agents like Claude Code, Cursor, and Codex, records them into a tamper resistant audit trail, and fires alerts when it detects dangerous behavioral sequences. The entire system runs locally on your machine with no cloud calls, so sensitive code and credentials never leave your environment.

The problem it addresses is straightforward. AI coding agents can now read files, execute commands, make network requests, and modify your codebase autonomously. That power is useful, but it also means a compromised or misbehaving agent could exfiltrate secrets, install backdoors, or push code you never reviewed. Traditional endpoint detection tools were not designed with agent session context in mind, so they miss patterns like credential access followed by network egress within the same task. Agentmetry fills that gap by understanding the agent lifecycle and correlating events across a session.

Under the hood, Agentmetry hooks into your IDE through two capture paths. Tier A wraps Model Context Protocol servers with an audit proxy that logs every tool call. Tier B uses native lifecycle hooks for supported agents to intercept actions before they execute. Each event gets normalized into a canonical schema, enriched with MITRE ATT&CK technique mappings, hashed to protect sensitive arguments, and appended to a local JSONL file with tamper evident hash chaining. A sequence detection engine runs over these events in real time, checking for predefined attack patterns like PR merged without review or outbound network call after file read.

The tool is aimed at SOC and DevSecOps teams who need visibility into what AI agents actually do, platform engineers who want incident evidence trails, and organizations with compliance requirements that call for auditable process documentation. If you are working in a regulated environment or simply do not trust third party agent code, this gives you the observability layer to answer what happened and when.

What sets Agentmetry apart from generic logging is the focus on sequence detection. It does not just record individual tool calls, it correlates them across the session and flags multi step attack chains. It also integrates with enterprise SIEM platforms like Elastic, Splunk, and Loki so security teams can fold agent activity into their existing dashboards. A real time web dashboard visualizes event streams, MITRE tactic breakdowns, and detection triage status if you prefer a standalone view.

The project is open source under the Apache 2.0 license. Installation is a git clone and a Python pip install, or a PowerShell script on Windows. There are no SaaS fees, no cloud egress billing, and all data stays on your hardware. The data footprint is around fifteen megabytes per developer per month based on median event volume. It is currently in public alpha, so expect rough edges and documentation gaps, but the core capture and detection logic is functional. Anyone managing AI agent usage at scale should consider adding this to their observability stack.