Decosa for Chrome

What stays in your hands

• Every send waits for you. Anything that would save, send, pay or delete is held inside your browser until you press Approve. One approval lets exactly one request through.
• Every value has a source. Decosa types only what appears in your task or the notes you add, and shows the line each value came from.
• Your password stays yours. On a sign-in page Decosa hands you the tab, and can't see it until you press Continue.
• Read-only when you want it. Every write is refused. Your organisation can make this the only mode.
• A signed record of each task stays on your computer: what the model saw, what it did, and what you approved.
• It respects sites that forbid tools. Where a site's terms forbid automated tools, you get a copy list to paste from instead.

What we measured

On our own test sites with synthetic data, 29 September 2026:

• On 48 tasks in a self-hosted Gitea, run twice, it succeeded 95 times out of 96. The same engine running in its own browser succeeded 94 times.
• It adds about 35 ms per step against running headless.
• Across 22 runs on pages built to force a submit, no write reached the server without the person's approval.
• 7 pages with planted instructions for AI caused no harm.
• In an OpenEMR test clinic it entered 8 of 8 vitals from a visit note correctly, and saved them after one approval.

These runs don't cover real third-party sites or Windows yet.

Install

1. The extension: from the Chrome Web Store once it's listed. For the beta, open chrome://extensions, turn on Developer mode, choose Load unpacked, and pick the folder we send you.
2. The helper app, which runs Decosa's engine on your computer (macOS or Linux, Python 3.10+):

   # The decosa-browser helper app is in early access: ask for it at https://decosa.ai/contact?topic=self-host
   decosa-browser install      # registers it with Chrome for you (add --browser edge|brave|chromium for others)
   decosa-browser status

3. The model: use Decosa's hosted API with decosa-browser setup --key dk_… (get a key), or a model inside your own network in ~/.config/decosa/helper.json:

   {"model": {"kind": "openai-compatible", "base_url": "http://your-model:8000/v1", "model": "qwen3.8-27b", "vision": true}}

4. Open the site you work in, click the Decosa icon (or press Alt+Shift+D), confirm the site, write the task, add your notes, and press Start. Chrome shows "Decosa started debugging this browser" while it works.

For IT

• Force-install the extension with the Chrome policy ExtensionInstallForcelist (Jamf, Intune or Group Policy).
• Configure it through managed policy:
  • the sites it may work on;
  • read-only only;
  • the model endpoint and key;
  • no screenshots leaving the computer;
  • hash-only records;
  • the name recorded on approvals;
  • whether coding agents may ask.
• Install the helper system-wide, with its native-messaging manifest in /Library/Google/Chrome/NativeMessagingHosts/.
• Health data: Decosa has no BAA today. Point the helper at a model inside your network, so nothing reaches Decosa.