OpenClaw 9.3 released: Browser automation can finally be watched in real time, and the AI operation web page is live broadcast

After waiting for a long time, this feature is finally here. OpenClaw v2026.9.3 is officially released, and the biggest highlight is just one sentence: browser automation can finally be viewed in real time.

In the past, AI operated the browser for you, and you could only look at the results - where it clicked, where it jumped, and where it got stuck, all relying on second-guessing. Now open the browser panel and you can watch every action of the AI just like watching a live broadcast.

What exactly has been changed in this update?

The official summary of this edition is six key words:Clean update recovery, faster session reconnection, real-time browser automation, revocable chat links, searchable meeting transcripts, cloud warehouse work. The release was not small - 1,844 PRs, 40 direct commits, 190 contributors.

The focus is on the browser. The new version of the browser panel canWatch Agent browser tabs redraw in real time, watch it open the page, click, enter, and jump all the way. If streaming is unavailable, it will automatically fall back to screenshot mode without a blank screen.

How to see it in real-time browser

The bottom layer is a real-time streaming channel, and the mechanism is very clear:

one-time token: Apply for a streaming token, 48-digit hexadecimal, valid within 60 seconds, can only be used once, will be rejected directly if it expires or is replayed.
WebSocket push: After connecting to the endpoint, the browser screen continues to be pushed over in WebSocket frames.
About 20 frames/second: Chrome sends one JPEG frame each time it is redrawn, and the rhythm is controlled at about 20 frames per second.
Lose frames if it’s slow: When the viewing end cannot keep up, the old frame will be skipped directly without queuing - ensuring that what you see is always the latest picture.

In addition to the picture, each frame also contains URL, viewport width and height, scroll position, and timestamp. So you not only see the picture, but also know exactly what state the page is in at the moment.

Security design is unambiguous

For things like real-time images, permissions must be blocked:

The token is bound to the gateway connection that initiates the request - once the connection is disconnected, the unused token is immediately invalidated, and the viewing terminal being watched is also closed.
A revoked connection will be "fence-fenced": the token cannot be upgraded to a stream, and the viewer cannot receive subsequent frames.
The current capture session is immediately ended when the navigation jumps, and the new page will only be restarted after the address verification passes - the residual frames of the previous page cannot enter the new stream.
Navigation rejected by the policy directly cuts off the stream

To put it simply, "being able to see" and "being able to see casually" are two different things. Address verification, navigation strategy, and connection binding are all included.

There’s another surprise for Mac users

The Mac app has a new native browser tab: external links will be opened in the same browser panel as native WebKit tabs, supporting navigation and one-time annotate/inspect screenshots. Moreover, these tabs belong to their own windows and will not be closed when switching conversations.

There are also several practical updates: Skill Workshop can now retain the skills learned by each agent across workspaces; Models settings have added independent account control and account sorting; chat links can be revoked; meeting records support search; and it also supports running warehouse-based work in the cloud.

Note before upgrading

Node version requirements: Requires Node 24.16.0 or above (24.x line), or Node 26.1.0 or above, Node 26 is officially recommended
Updated recovery is cleaner: Can be restored more smoothly after interruption, and the session can be reconnected faster

Why is this important?

The biggest trust issue with AI Agent is "black box execution". You ask it to do something, and it replies "It's done" to you, but you don't know what it did in the middle. The real-time picture opens the black box a little bit - at least, you can see it.

From "looking at the results" to "looking at the process", this is a key step for browser automation to become usable and trustworthy. Only when automation is visible can we dare to let it do real work.

AI operates the browser and is no longer a black box.

Reference: OpenClaw official release notes v2026.9.3, browser control API documentation