Every time you ask an AI for something, a lot happens behind the scenes that you never see. Searches run. Scripts execute. Files get read. Trace records all of it, seals it so nobody can alter it later, and lets you save the parts that worked so you never have to figure them out twice.
Each step gets a colored bar on the left edge. Green means it ran and the result checks out. Amber means it ran but nobody verified it. Red means it broke. You can scan a whole session in seconds.
It ran, and the result is confirmed. You can rely on it.
It ran, but the result hasn't been checked yet.
It failed, got blocked, or didn't have what it needed to run.
It's still running. Watch it finish in real time.
Whether you use a hook, a proxy, or a browser extension, everything Trace captures ends up in the same format. That means records from different tools can be compared and checked against each other.
Hooks fire after a tool finishes and record exactly what went in and what came out. The model never touches this data. It is the most trustworthy source, observed:hook.
A small proxy sits between your app and the AI provider and records the traffic. You change one setting and you are done, observed:proxy.
A browser extension watches the conversation as it appears on screen. The easiest way to start, observed:extension.
If you run more than one channel at once, Trace compares what each one saw. When they disagree, it tells you plainly instead of quietly picking one.
Pin any step and it is yours forever, stored at an address based on its own fingerprint. Bring it into a new session months later and you get back exactly what happened the first time. Not a summary. The real thing, verified.
If the AI says "I searched the web," that is a claim, not proof. Trace only calls something observed when a hook, proxy, or extension actually saw it happen. A claim can never be promoted to proof.
| Tag | Source | Trust model |
|---|---|---|
| observed:hook | PostToolUse hook | The real thing. Nothing ranks above it. |
| observed:proxy | API stream interception | Very accurate, with full context. |
| observed:extension | Browser DOM / network | Screen-level detail. Good, not perfect. |
| reported:model | Model's self-assertion | A claim, not proof. Stays that way. |
| derived:transcript | Post-hoc transcript parsing | Reconstructed after the fact. Fills gaps. |
Put the proxy in front of your company's AI traffic and every interaction is recorded with a tamper-proof trail. When an auditor or regulator asks what happened, you don't reconstruct it from memory. You show them the record, and the math backs you up.