A neutral layer for every AI agent.

Where it came from

trackline grew out of the evaluation harness for Lira Intelligence, a production AI support agent with retrieval over customer knowledge bases and risk-tiered tool calling. The principles it carries came from there: everything that can be checked by code is checked by code, safety rules get no tolerance, and a check that errors fails loudly instead of passing quietly.

Why neutral

A vendor can govern its own agent. Only something that belongs to none of them can govern all of them the same way. trackline is neutral by construction, not by promise: the engine is open source, it runs on your machine, it treats Claude Code, Codex, Cursor and any MCP client alike, and no vendor's model sits in the path. Without an account it sends nothing anywhere.

How it is built

Every claim on this site was measured before it was built on, and the 8 experiments are published with their method and limits. The part that runs before every action is written in Go, because it runs on every tool call: starting Node cost 88 ms a call against 6.5 ms for Go, and the whole check runs in 11 to 14 ms.

Who builds it

trackline is built by Yerins Abraham. The longer story of the problem and the research behind it is in Nothing notices when an agent drifts.

Questions, or running it across a team: hello@trackline.dev