Ghost.exe presents

Screener

See what you're about to tell an AI, before it does.

… …

Under the hood, live

Vercel UI SDK builds the page, and lets agents use it

Components on this page
0 of 0 kinds
Agent steps run
0
Steps held for a person
0 by data-agent-confirm

Deliverd screens it, and asks a person

Checks in this tab
0 at 0 ms each
Values swapped for placeholders
0
People asked
0 0 yes · 0 no

Your prompt

Try one:
Prompt Checked as you type, in this tab. Nothing leaves your machine until you press Send.

Drop text files here to check them too.

Notes, CSVs, logs, code, .env files. Up to 1 MB each.

What the model may receive

You can change this yourself. An agent has to ask a person first.

Highest classification Your own confidential words Screener looks for these too. They stay in this tab and on your machine.

Verdict

Waiting for something to check

      Screener doesn't look for people's names yet. Patterns can't find them; a model can. How

      What the model will see

      
                

      Watch an agent try

      An agent drives this page through Vercel UI SDK's agent bridge. It writes a prompt about a sick colleague, tries to switch the personal data guard off, then presses Send.

        What have you already pasted?

        ChatGPT and Claude both let you download everything you've ever said to them. Drop that export here and Screener counts the email addresses, card numbers, keys and the rest that went in, and in which conversations. The file is read in this tab and never uploaded. It shows kinds and counts, never the values.

        • ChatGPT: Settings → Data controls → Export data. The link arrives by email.
        • Claude: Settings → Privacy → Export data.

        Drop your export here (the .zip, or conversations.json)

        How Screener works

        1. It looks, in your browser. As you type, Deliverd Screen looks for email addresses, phone numbers, card numbers, IBANs, national ID numbers, keys, passwords, special-category words such as a diagnosis, and your own confidential words. It runs in this tab, with no network call.
        2. It decides again when you press Send. Run it yourself and the little server that comes with it checks the same text again; it never takes the page's word for it. In the hosted demo there's no server, so the same check runs in your tab.
        3. Four answers. Allow sends it as it is. Redact swaps what the model may not receive for placeholders such as [EMAIL_1], and puts the real values back in the answer. Ask puts it to a person first. Deny sends nothing: secrets never go.
        4. A person decides the hard ones. In the playground, that's you, at the approver's phone on this page. Connect Deliverd and it's the people you name, in Slack, Teams or on their phone, with a record of who said yes. They're told what kind of data it is, never the data.
        5. Agents ask too. The page is built with Vercel UI SDK, so an agent can use it. The model settings carry a data-agent-confirm mark, so an agent that tries to loosen them waits for a person. A browser's own agent can use the page through WebMCP, under the same rule.

        A net, not a certificate. Structured identifiers are found well; names are not found at all without a model. Text with placeholders in it can still be personal data. Use it to send less, never as proof a prompt was clean.