A browser that runs on your own terms.
Give your agent a working browser inside restricted environments, with a receipt you can audit for every action.
What it is: one browser, four interfaces
-
Pure HTML relay
Plain HTML forms in, a rendered receipt back, with scripts switched off.
-
File dialog and download interceptor
Answers file pickers from your workspace and saves downloads there.
-
Pixel coordinate actuator
Acts by screen position when a page hides its structure.
-
Sovereign virtual desktop
Runs browser sessions in a virtual desktop, on your own terms.
Who it is for
- Agents in locked-down webviews
- Agents that run inside strict webviews or sandboxes, such as Muse and Dots, with scripts switched off.
- Jobs that move files
- Work that uploads a document or collects a download as one of its steps.
- Pages that are hard to read
- Sites with obfuscated selectors, where a position on the screen is the dependable handle.
- People who audit agent work
- Anyone who needs a record of each action an agent took in a browser.
How you stay in control
A receipt for every action
The server renders each result and returns it as a receipt, like the example above, so every step your agent took is on the record.
Files by content address
Uploads and downloads go through a content-addressed workspace. Each file is named by its content, so you can check that what arrived is what was sent.
Actions checked on screen
Each pointer and keyboard action carries coordinate verification from Qwen3.5-VL and UI-TARS, grounded in what the screen shows.
Setup
Start with the interface that fits where your agent runs, then add others as the job grows.
| Where your agent runs | Start with | Have ready |
|---|---|---|
| A webview or sandbox with scripts switched off | Pure HTML relay | A way for your agent to send an HTTP POST form |
| A Chrome session your agent drives | File dialog and download interceptor | A Chrome DevTools Protocol connection and a workspace folder |
| Pages with obfuscated selectors | Pixel coordinate actuator | Screenshots of the page and access to Qwen3.5-VL or UI-TARS |
| Work that spans a whole desktop session | Sovereign virtual desktop | A host for the desktop session, on your terms |
Questions
What is a pure HTML relay?
A relay your agent drives with plain HTML forms. It posts a form, the server acts, and the reply is a rendered page with the receipt.
Which sandboxes does it work in?
Strict webviews and sandboxes that run with scripts switched off, including Muse and Dots.
How are file pickers handled?
Through the Chrome DevTools Protocol, with Page.setInterceptFileChooserDialog. The browser answers each file picker from a content-addressed workspace, and downloads stream into the same workspace.
How do I check what it did?
Read the receipts. Each action has one, rendered on the server, and each file carries the hash that names it.
See a whole job run, step by step.
The sandbox walks through three example jobs, each step checked before the next.