onetrace · open source · runs on your machine

Your AI changed. Find the first step that did.

onetrace records each step of your pipeline as a receipt anyone can check. Compare two runs and it points at the first step whose output differs, and what was recorded there.

Install
pip install onetrace "onetrace-console[neighbours]" "onetrace-explain[keyring]"
A comparison on the live demo: five steps, with classify marked First difference, the two steps after it marked Changed after it, and a table of what differed at each step.
A comparison on the live demo. The first difference is at classify, where a rule setting changed.
What it does

Record, compare, prove.

01

Record every step

One receipt per step: what went in, what came out, the tool and its settings. Add it with your coding agent, with no code change, or step by step.

02

Find the first difference

Open two runs in the console. It names the first step whose output differs, the settings that changed there, and the exact text.

03

Check without trusting the code

A separate verifier reads only the bytes on disk. Change a receipt after the run and it says which one.

Debugging an agent

Your invoice agent paid €4,000 on a €40 invoice. Where do you start?

Check it was recorded. Read the bad run. Compare it with a good one. Find runs like it. Ask what changed, and whether it's just noise. Re-run it without paying twice. Check who allowed it. Ask for an explanation you can check. Nine avenues, one record they all read from.

Walk through the nine avenues →

An illustrative example, not a real event.

yesterdayintake→extract→lookup→approve→pay
todayintake→extract→lookup→approve→pay

The wrong payment shows at pay. onetrace points at extract, the first step whose output differs.

Explain

Facts first. Then a model's reading, labelled as one.

onetrace-explain takes the two runs and separates what the record shows from what a model suggests.

recordedWhat changed and the evidence. No model, no network calls.
AI-generatedWhat may have happened. Only with a model you configure, and every claim must cite the record or it's dropped.
How explain works →
The live demo's split comparison: the short answer at the top, then the kept explanation, labelled as written by a model.
How it stays honest

Local, checkable, and clear about its limits.

No account

Runs are folders on your disk. Nothing is sent anywhere unless you ask: Explain with a model you configure, or an outside timestamp.

Independent check

onetrace-verify uses the Python standard library only and trusts none of the code that wrote the record.

Ready for CI

onetrace diff exits 0 when two runs match and 1 when they diverge, so a job can gate on it.

Stated limits

A faithfully recorded wrong answer still verifies. What it doesn't prove.

See it on real runs first.

The live demo holds 58 runs of 11 example pipelines, some broken on purpose. Or run the same demo locally with one command.

Run the demo locally
onetrace-console demo --open