Try Aetius: public preview build
Two commands, then your own traces.
The same build we run: installed with pipx, configured in one screen, pointed at the trace data you already collect.
$ curl -fsSL https://aetius-webpage.vercel.app/downloads/install_latest.sh | bash -s -- https://aetius-webpage.vercel.app/downloads/aetius-latest.whl$ aetiusThe script fetches the wheel from this site, renames it to the spec-compliant filename pipx requires, and installs it with pipx. Nothing else touches your system.
Before you run it
Anthropic API key
Runs the attack engine and the reconstruction models. Any key with API access works.
Langfuse or Braintrust key
Read access to the project whose traces you want to test against. One active trace source at a time; you can switch sources later from Settings.
Python 3.11+ and pipx
The only runtime requirements. The install script downloads the wheel and installs it with pipx, so re-running it upgrades in place.
What happens next
Credentials screen
First launch opens a three-step credentials flow: your Anthropic key, a trace-source picker, then that source's fields. Keys already set as environment variables are detected and skipped. Each key is validated against its service, then written to a local env file on your machine. No key is ever rendered on screen or sent anywhere except the service itself.
Straight into your traces
The flow lands you directly in Add Environment. Point it at a project, and it pulls traces and reconstructs your agent's structure: roles, tools, handoffs.
Run a comparison
Pick a baseline and a candidate. The run is sized before it's paid for, so you see the run count and what it can detect up front. It ends in a verdict: FLAGGED, no difference detected, or INCONCLUSIVE with how much more evidence it would take.