Shft Agent

How Shft keeps you safe

An assistant that can do things for you has to be careful by design. Here is what Shft Agent does about that, in plain words. For the legal detail, see the privacy notice and the terms.

You decide how much it does on its own

By default, Shft asks before sending. You control its autonomy. Buying and deleting always require approval.

• Ask me first is the default. Before anything goes to another person, you see what it is and who it's for, and you say yes or no.
• You can give it more freedom. In Autonomous mode it can send on its own, and a standing rule can let one kind of thing go without asking, like invites to people at your own company. You can see and remove every rule on your Access page.
• Buying and deleting always wait for your OK, in every mode. No setting or rule changes that.
• After it reads outside content, it asks again. Once a job has read a web page, an email or a file from someone else, sending waits for you whatever the mode.
• Looking things up and making files are things it does on its own. Researching, reading your documents and making a PDF, Word, Excel or PowerPoint file doesn't send anything anywhere.
• A filled-in form comes back to you. It fills in the PDF forms you send and hands you the copy to check. It never sends it for you.
• When it emails someone, it does so from its own address and says it's an AI agent. By default, it asks you first.

Text from outside is information, not instructions

Websites, emails and files can say anything, including things written to trick an assistant. Shft Agent treats what it reads as information to work from, never as instructions to follow. Only you tell it what to do.

Your passwords stay yours

• You do the sign-in. When a site needs a sign-in or a code, it hands you its browser for that step, and carries on when you give it back.
• Saved logins work like a sealed vault. It can sign in to a site with one, but it never sees the password.
• Banks, payments and government sites are off limits. It refuses them.
• You can watch, and take over at any time. You can see its browser live, open it full size beside the chat, and take the controls whenever you like.

It only reaches what you've given it

• Calendars are read only. You paste the private calendar address from Google, Apple or Outlook. It books on its own calendar, and finds free times with other people's agents.
• Google Drive is the files you pick, and only those, read only.
• GitHub is public repositories and the private ones you choose, read only. It proposes changes as patches for you to look over.
• Slack, Notion, Linear, Todoist, Asana, HubSpot and similar apps: it reads freely and, by default, asks before it sends or changes anything.
• Apps on your Mac: it asks which apps it may use for a job, anything that sends, buys or deletes goes to you first, and a Stop bar is always on screen.
• Coding on your Mac happens in the project folders you allow, in a sandbox with no internet unless you OK it. It can't see your keys or your other folders, and every change is shown as a diff you can undo.
• Python for number work runs in a sandbox of its own.

Your context stays with Shft

• Your context and memory stay with Shft, not with the model companies.
• Each step sends a model only what that step needs.
• Your data isn't used to train models.
• It's stored in Frankfurt, Germany, and we don't sell it or use it for ads.
• Its memory is yours to read. You can see what it remembers about you, and edit it or clear it.

A record of what it did

Every job gets a receipt. There's a private activity log, and an Access page that lists everything your agent can get into, so you can check it whenever you like.

Honest limits

Shft Agent is new, and AI can get things wrong. That's why, by default, it shows you what it's about to do before it sends anything, why it always asks before buying or deleting, and why every job leaves a receipt. If something looks off, say no. If you see something that doesn't look right, tell us at hello@meetshft.com.