What Keegi holds, and how to make it stop.

Keegi is two products sharing one engine, and they hold different things about different people. A company workspace reads work its admin connected. Keegi Public holds only what a person typed about themselves. Nothing crosses between them. They are separate tenants, and the database enforces that, not application code.

What Keegi holds

• Your account. An email address. If you signed in with Google, the account identifier Google returns. There is no password anywhere in Keegi, so there is none to store. Sessions are cookies.
• In a company workspace, your work. Whatever the connectors your admin switched on can read: tickets, comments, documents, messages in the channels they included. Keegi turns that into one capability card and keeps the citations behind it. It never keeps a productivity score, an activity count, or anything you could paste into a performance review. That boundary is in the architecture, not in a settings page. Security draws it exactly.
• In Keegi Public, what you wrote. The headline, the description and the topics you typed on your own card, plus public metadata from any GitHub repositories or domains you connected yourself. Nothing else. Keegi Public holds no employer data.
• The questions people ask. Every question is logged with the workspace it was asked in and the account that asked it. That log is what the rate limits count and what the gaps dashboard is built from. Your own card shows you who has asked about you.
• An address you left before there was an account. The waitlist takes an email. The five-question page takes an email and the answers you gave about the shape of the company. Neither is a score. We keep them so we can write with a next step. The page cannot be used to ask whether someone else is already on that list.

Where it runs

The database is in Ireland and every server function runs in Dublin, in the same AWS region. That is pinned in configuration, not asserted here. One thing leaves that envelope: the model. A question and the evidence text behind an answer go to whichever provider can serve it, and we do not control where that is. The provider deletes both once it has answered.

We tried to fix that and could not, which is worth saying plainly. For four days in August we routed every model call to a European data centre. It broke the part of Keegi that writes the answer, on both of the two providers that offer Europe, so we took it out. The switch is still in the code and goes back on the day one of them works.

Who else sees it

The database, the runtime, the model gateway and the providers behind it, always. Jira, Linear, Slack and Notion only if your admin connected them, and a workspace that never connects one sends it nothing. Resend delivers relayed messages. The full list, with what each one actually receives, was assembled by reading every outbound host in the codebase, so it says what the software sends rather than what we remember it sending.

Sentry gets told when the server breaks. It is set up to receive the stack trace and nothing else: not the question you asked, not the text Keegi read about anybody, not the rows that came back from the database, not your cookies. Those are all things it collects by default and they are all switched off, and a check in our build fails if any of them comes back on.

There is no analytics script, no session recording, no tag manager and no advertising pixel on any page. Not disabled by default. There is no third-party script to disable.

If somebody sends you a message through Keegi

You will get one email. It says who is asking and what they want, and it does not carry their address. Replying is what discloses yours to them, and the email says so before you reply. Keegi will not send you a second one about the same question. The link at the foot of it stops all of them, and it needs no account and no reason.

What you can do

• See everything Keegi holds about you, as it holds it. Your card is a page you open, not an export you request.
• Switch off any single piece of evidence, or stop being findable altogether. One click, reversible, no reason required.
• Ask for the lot to be deleted. Write to the address below.
• Refuse a relayed message. Every one carries a link that stops the next one, and the mail client’s own one-click unsubscribe works too.

What is not settled yet

Four things here need a lawyer, and they are not written yet: the legal entity and its address, the lawful basis for each kind of processing, how long each thing is kept, and the transfer mechanism for model inference. Our own documents mark them open too. If one of them is the answer you need in order to decide, ask. You will get the real one instead of a paragraph.

Questions, deletion requests, or anything on the list above: someone@mail.keegi.xyz. See also Security for the boundary Keegi does not cross, and Terms.