A← Index

offstage

A screen of its own for each AI agent, so my desktop stays mine.

RoleMaintainer of a fork
TypeDeveloper tool · MCP server
StackGo · cage · Bare · MCP
StatusBeta · in daily use

IThe problem

When a coding agent opens a browser on my desktop, it can cover my work, move the cursor or take focus. I wanted agents to test interfaces while I keep using the machine.

offstage began as a fork of an open-source project that gives an agent a separate, in-memory screen through the cage compositor. I maintain it and wrote most of its current code: text snapshots, actions with verification, recording, task tools, the Bare browser integration, the beta report pipeline and self-update.

354of 368 commits are mine
3text sources: CDP, AT-SPI, OCR
4screens per session by default

IIHow it fits together

Each screen is its own compositor with its own output and seat. My desktop never learns that the screen exists, so my monitors, workspaces, focus and cursor never move.

IIIRead before taking a screenshot

A screenshot costs about 1,300 tokens. Most actions only need a control's name, role and position, so the agent reads the screen as text first.

snapshot
Lists the elements of the screen with a ref, a role, a name and coordinates. The text comes from the DevTools protocol, then AT-SPI, then OCR.
act and find
act clicks, types or scrolls on a ref and returns what changed. find waits for an element instead of a blind sleep.
A verdict for each action
The server compares a frame before and after each action: CONFIRMED, PARTIAL, DIVERGED or NO_OP. A click that did nothing does not pass silently.
Screenshots when needed
For a canvas or a layout question, the agent still takes a screenshot. It can write the image to a file, so the image stays out of its context.

IVTest and show the result

Bare, my own browser
The bundled browser is Bare. The server exposes its tools as browser_*. Chrome is not required.
Recording
Records crop to the page viewport and cut idle time. The pointer glides to each target and shows a ring at each click, so a demo is easy to follow.
Saved logins
browser_save_state and browser_ensure_session keep a login, so an end-to-end test starts signed in.
Agents file their own reports
In the closed beta, report_issue sends a report with a replayable trace. The server updates itself about once an hour, so fixes ship after release.

Limits. A separate screen stops interruptions. It does not isolate apps from the file system or the network. Each screen uses memory, so the number per session is capped. Accessibility coverage varies by app.

Windows. Windows runs Bare headless. It opens no window, so my desktop is never touched. It needs Windows 11 and no Chrome. It has a smaller tool set: the browser tools and report_issue, with no native apps, desktop tools or recording.

SignedArjun