What we never collect, and what we will never build
The private channel, what you say to your agent, is protected by schema, not by policy. What is searchable is only what someone deliberately composed into a shared record: the same exposure as a commit message.
Never leaves your machine
- Prompts and transcripts
- Tool output and terminal output
- Source code and diffs
- Plans and scratch files
- Anything an agent read or wrote on disk
What travels
- Session id, machine name, worktree path, entrypoint
- Event kinds and timestamps
- Refs: work items, PR numbers, branch names, decision ids
- Titles, notes and messages you or your agent deliberately authored, secret-scanned before send
- Hashes where a fact needs an identity
Known seam: a session title may be model-generated from session content. It is secret-scanned before it is sent, and there is no switch to turn it off yet.
Will not build
- Per-person productivity or hours dashboards
- Rankings or league tables of any metric
- Sentiment analysis over anything anyone wrote
- Exports whose grain is the person, not the work
- A hosted raw-capture tier, at any price.
The waitlist stores your email address, your answer and when you signed up (so invites go out in order) on Cloudflare, and uses them only to invite you to the beta.
This is the promise, not the policy. Legal pages come before public sign-up; the alpha is private.