OpenComputer 0.5.0 — actions that check themselves, Take over, and a browser of the agent's own

Every action now says what changed and whether it worked, you can stop the agent or approve what matters from the activity card, and agents can browse in a Chrome of their own.

Watch the 0.5.0 demo, recorded unedited by this release's workflow on a GitHub macOS runner: an agent fills in and submits a background app while the terminal keeps the foreground, checks its own results, asks before a sensitive press, is stopped with Take over, and is refused in the terminal it runs in.

Actions that check themselves

• Diffs by default. After an action the server waits until the app's own accessibility notifications go quiet, then returns only what was added, removed, or changed. Element IDs stay valid, and "no change" now tells an action with no effect apart from one the outline can't show.
• expect and verify. Say what should happen; the result is confirmed, deviation, or unverifiable. verify checks without acting.
• Shortcuts in background apps. pressKey with via: "menu" presses the menu item that owns the shortcut; menu runs a command by path, such as File > Export as PDF….
• See what the outline can't. Screenshot marks (element IDs drawn on the picture), region zoom, on-device ocr, and maxEdge; readText for long text; waitFor; batch for up to 20 calls.
• Errors say what to do next: retrySafe, plus next and hint.
• Access levels. bundle:read and bundle:click alongside full access, and under * the terminal, editor, or chat app running the agent is read-only.
• Checkable record. release summarizes the session; the action log is hash-chained (opencomputer log --verify, --summary).

See, stop, and approve

• Activity card. The dots open on hover into a card with the agent, app, window, elapsed time, and the last three steps.
• Take over. Pauses the session until you press Resume; actions answer PAUSED_BY_USER with nothing sent.
• Approvals. OPENCOMPUTER_CONFIRM=sensitive (or a bundle ID, or all) asks on the card before presses like Send, Pay, or Delete. Allow takes a held press and is never offered to automation.
• Named ghost cursor with the agent's name tag.

Agents browse in their own Chrome

• open with app: "agent-browser" starts a second Chrome with its own profile, in the background, so agent tabs never land in your windows. com.google.Chrome always means your Chrome.
• open on the agent browser answers USER_BUSY instead of starting Chrome while a fullscreen app or game is in front.
• Screenshots fall back to the system screencapture tool when ScreenCaptureKit stalls.

Known limits: starting the agent browser can bring Chrome forward for a moment before the focus guard hands the foreground back. The demo's Take over step moves the real pointer to play the person; OpenComputer itself never posts mouse events.

Validation: this release's workflow ran the unit tests, the live fixture suites (25 and 19 assertions), and the demo on a GitHub macOS runner before publishing; the live overlay UI test passes in the Overlay screenshots workflow. The agent browser is not covered by the runner tests.

Requires macOS 14+ and Node 22+. The runtime ZIP contains an ad-hoc signed universal Apple Silicon/Intel helper. Reconnect clients after updating. Full details are in CHANGELOG.md.