ATTRIBUT captures metadata only. Your prompts, the model's responses, your code, and your diffs never leave your machine.
The principle
The capture hook reads your AI coding sessions and sends structured metadata so your work can be attributed and measured. It never sends the content of your work. There is no denylist to trust: if a field isn't in the contract, it never leaves your machine, and a test enforces that.
What is captured
Captured | Detail |
Session id | Identifies one chat session |
| Random, stable per machine — see below |
Model | Primary model id used |
Token counts | Input, output, and cache tokens |
Timing | Session start, end, and duration |
| Working-directory path and git branch |
Commit SHAs | From |
Counts | Turn count and tool-call count |
Tool use | Tool NAMES and how many times each ran |
Line changes | Structural counts only — numbers, no content |
The one content exception: title
title is your chat's short title — the model-generated (or custom) label for the session — capped at 200 characters. It is the only field derived from content, added by an explicit product decision. Nothing else about what you typed or the model returned is sent.
Never sent
Prompt text
Assistant responses
File or diff contents
Tool input arguments
PR and commit message bodies
About device_uuid
device_uuid is a random value that is stable per machine. It is not your identity — you are resolved server-side from your token. It only groups activity by device.
About repo
repo is the absolute working-directory path, so it includes your OS username and folder layout. It's used server-side as the repo's identity.
How the data travels
The metadata is gzipped and sent over HTTPS with a bearer token. Your ingest token lives in an owner-only (0600) file, never in a settings file and never on the command line.
Why you can trust this
ATTRIBUT's CLI is source-available under PolyForm Shield 1.0.0 — the full source is public at github.com/attribut-ai/cli, so anyone can read exactly what is and isn't sent. You don't have to take our word for it: inspect the code, or prove it on your own machine.
Related
