One command

Run any agent harness. Record every model call.

Give it a GitHub repo that holds an AI agent harness and a task. It downloads the repo, has an AI write a Docker sandbox for it, builds the sandbox, puts a recording proxy between the harness and the model, and runs the task.

Every model call the harness makes is captured on the wire, in one format, whatever the harness is. The harness never holds credentials, and nothing in it is modified.

Not public yet. If you build harnesses, or measure them, we would like to hear from you.