BlackboxTime-travel debugger for AI agents

Record a run of your own Anthropic or OpenAI agent. Replay it offline, fork it with one tool result changed, and diff to the first step where the runs part.

Quickstart

See it find a root cause

How it works

GitHub

Record your own agent

Pass bb.fetch to the official SDK client and wrap your tools. Model calls and tool results go into a hash-chained JSON cassette.

Replay with no model calls

Your real code runs again against the cassette. Each request is compared with the recorded one, and the first difference is named. No API key needed.

Fork and diff

Hand the agent a different tool result at any recorded step, continue live or from scripted replies, and see the first divergence and the new answer.

Pin it in CI

blackbox assert and blackbox replay turn a committed cassette into an offline regression test.

blackbox inspect shows a telemetry reading captured before the fault; blackbox diff shows the first divergence at step 5 and the work order changing from routine to urgent

The recording above is from the fleet-triage example: a maintenance agent kept an overheating van in service because its telemetry tool served a stale reading. Blackbox reproduces the run, isolates the reading, tests the fix by forking, and pins it.

Blackbox is unrelated to Blackbox AI, the coding assistant.