Your data, plainly.
We don't train on your work, we don't delete it on a timer, and you can take all of it with you. Here are the details — including the parts still being tightened.
Not trained on by us
Provider terms differ, and we name each one's below
Signed at the write path
Every change carries its author — enforced, not conventional
No retention timer
Nothing expires on a schedule; trash holds until you empty it
Exports as plain git
Full revision history as real commits, walkable offline
Training
We never use your workspace content to train any model of ours. We don't operate models at all — we call other companies' APIs. Under Anthropic's, OpenAI's and xAI's published API terms, and Google's for the paid service we use, content sent that way isn't used for training by default. DeepSeek's terms don't rule it out, so a chat on DeepSeek is your choice for what you put in it, and isn't offered your Google connections. To be precise about the mechanism: we rely on those standard terms, not on a separately negotiated contract of our own. If that distinction matters to you, read their terms directly — we'd rather point you there than paraphrase them.
Who can receive your content
When you ask an AI to work, the files it needs go to the provider running that task. This is the complete list of third parties that can receive your content or personal data. If we add one, we'll update this page.
| Who | Why | What |
|---|---|---|
| Anthropic · OpenAI · Google · DeepSeek · xAI | Run the AI task you asked for | The files needed for that task, depending on the model chosen |
| OpenAI (search indexing) | Builds the index that makes your workspace searchable | File text, as files are written |
| Supabase | Database and authentication | Your workspace contents and account |
| Render · Vercel | Application hosting | Traffic in transit |
| Sentry | Crash and error reporting | Configured to collect no personal data |
| Resend | Transactional email | Recipients and message contents |
| Stripe | Payments | Billing details |
Deletion
Nothing expires on a schedule. Trash holds until you empty it — deliberately, because a timer means the system destroying your work with nobody watching. When you delete something permanently, or delete your account, removal is immediate rather than queued.
Access
Who can read a file is access you give and can take back — not a property of who or what is asking. Connected assistants reach your workspace through OAuth you approve, and every write they make is signed with their name. A connected assistant can read, write, move, delete, and share files on your behalf — the same reach you have — so connect ones you trust.
Your computer and your browser
The desktop app and the Chrome extension act on your own machine, and what they may touch is decided there. The folders you allow, the programs and sites you keep them out of, and the switch itself are stored on that machine, not in your workspace.
A file on your computer stays on your computer. What your agent reads from it goes to the provider running that piece of work, like anything else it reads. A Word, Excel or PowerPoint file passes through yarnnn while it is being read or changed, and is not kept. A picture of a window or a page, taken to check the work, is used for that one step and is not stored.
What we do keep is the record: which file or page was worked on, what was done, by whom and when. The extension asks Chrome for no permission to read your cookies, your clipboard or your browsing history, and what is typed in a password field never leaves the page.
What you can take
The whole workspace exports as a standard git repository, with the full revision history as real commits — readable with tools you already have, on a machine we have nothing to do with. Conversations aren't included yet.
It's in Workspace Settings, under Danger Zone — available any day, not just on the way out. The download names anything it couldn't include, so you always know what you have.
What's not done yet
On the roadmap. Two things we're tightening, both scheduled work rather than someday-maybe: some stored file contents can persist in backing storage after deletion, and reads of private file bodies still lean on application-layer checks rather than database-level rules. We name them here because you'd have no way to find them otherwise.
Not yet started. We hold no SOC 2 or ISO 27001 certification. Those are third-party audits, and we'd rather tell you we haven't done one than let a badge imply we have — we'll pursue them when our customers need them, and we'll say so here when that work begins. A DPA or BAA isn't an audit but a signed agreement: we don't offer one off the shelf today, so if you need either, talk to us and we'll tell you honestly where we stand.
The formal version is our privacy policy. Questions: contact@yarnnn.com.