For AI agents
Mask it. Prove it.
The privacy and proof layer for AI — for people and agents.
AI agents read your files, your terminal and your configs, and send them to an AI on their own — no person checks each request. Omna masks every one on your device before it leaves, and leaves a receipt for each.
$curl -fsSL https://omna.dev/cli/install.sh | bashA real Claude Code run through Omna on your computer, with fake data. Answer shortened.
Agents need it more than people do.
Whole files, .env files, logs and terminal output — exactly where keys and customer data live.
One task can be many requests to the AI, all sent on their own. Omna checks every one, every time.
Omna puts names and emails back in what the agent writes, so its edits still work. Keys and passwords stay as labels there, so a trick can't send them anywhere.
How it works with an agent.
The agent sends a request.
Omna catches it.
On your computer, before it leaves — with your own key or login, to the AI you already use.
Omna masks it.
The AI sees labels.
A key becomes [SECRET_AWS_KEY_1], an email [EMAIL_1]. Real values come back in the answer you read. In the agent's commands and files, only personal data comes back — never a key.
Every request gets a receipt.
You get the report.
Which agents ran, how many requests, and what was caught — counts only, never content.
Works with the agents you use.
Automatic
Connected by the install line, if they're on your computer.
One setting, typed once
Paste Omna's address into the tool's settings.
Building your own agent?
If it uses the Anthropic or OpenAI SDK, point it at Omna — one line. Everything else stays the same: your key, your model, your code.
Built to fail safe.
If Omna isn't running, agents can't connect at all. Nothing leaves unmasked.
Agents always get labels like [EMAIL_1], never fake values — a fake email could be committed by mistake; a label can't hide.
The only place a request goes is the AI you were already using. Receipts stay on your computer.