v2.3.8: Connections That Drop Now Heal Themselves, and Your Employees Can Talk

v2.3.8 is a reliability release, and it has one theme running through almost every line: connections that used to drop and stay dropped now heal themselves. The browser extension, the WhatsApp connector, the backend itself, all of them got better at coming back without you having to step in. Plus a new text-to-speech skill that every employee now ships with, and a handful of chat improvements.

The browser extension now heals itself

If you connect an employee to your Chrome browser, you may have hit this: the gateway drops (a deploy, a restart, a flaky network moment), the extension tries to reconnect for a few minutes, gives up, and then just sits there. The connection was back, but the extension had stopped looking, so your agent lost access to the browser until you manually reopened the extension. Sometimes the only fix was restarting the whole stack.

We rebuilt this on every layer.

The extension now reconnects forever instead of quitting after about three and a half minutes, and it keeps a small background timer that survives Chrome putting it to sleep, so a reconnect scheduled during an outage is no longer forgotten. On the server side, the gateway now actively checks each connection and clears out dead ones, and there was a particularly annoying bug where a single stale connection at the front of the line made every browser action fail even though a healthy connection was sitting right behind it. That is the root cause of "we had to restart the app for browsers to work again", and it is fixed.

We also added a watchdog that watches the backend's own pulse: if the backend stops responding, it gets automatically restarted. Previously a bad crash at startup could leave a whole node silently down, serving error pages while looking alive. Now it recovers on its own.

The upgraded extension (version 1.1.1) is already on the Chrome Web Store, so Chrome updates you automatically.

Unassign a browser in one click

Connecting a browser to an employee used to be a one-way door: once assigned, the agent kept driving Chrome and there was no clean way to cut it off from inside the product.

There is now an Unassign browser button right next to Assign. One click revokes the employee's browser access immediately and only for that employee (everyone else keeps theirs), and the agent is told it no longer has browser access so it stops trying. The chat shows a clear "Browser Automation Revoked" badge.

WhatsApp stops giving up

The WhatsApp connector had a similar problem. After a handful of failed reconnect attempts, which a brief network blip was enough to trigger, it gave up permanently and would not come back until the whole stack was restarted. And even a restart did not always revive it.

Your login was always saved correctly; the issue was purely in the reconnection logic. Now the connector retries indefinitely with sensible backoff, and it restores a wider range of accounts when the system boots, so an account that had dropped earlier comes back on its own. A genuinely logged-out session still stops cleanly and asks you to re-scan, as it should. Tested live: after a restart, the assigned account reconnects by itself, no QR re-scan needed.

Questions from your agent no longer freeze the chat

A recent Claude Code update introduced a fancy tabbed picker for when the agent wants to ask you a question or get approval on a plan. The catch: that picker did not register in our chat until you answered it, which meant the chat could deadlock, the message cut off and nothing to click. On custom-model employees the same questions came out as ugly raw code.

Both are fixed. Your agent now asks its questions and presents its plans in plain text in the chat, the way you would expect, on every kind of employee. Already-running employees pick this up after a session clear.

Your employees can talk now

There is a new text-to-speech skill, and it is installed by default on every new employee (along with the task scheduler). It uses Microsoft's neural voices: 322 voices across 75 languages, free, no API key needed. Each employee gets a sensible default voice picked from its profile and your language, and it can output a proper voice note, so your employees can reply with audio on WhatsApp or Telegram, not just text.

Chat and file improvements

• Tables look like tables. Markdown tables used to show up as raw text full of pipe characters. They now render as proper styled tables, and crucially they are readable in dark mode, which the previous attempt was not.
• Audio plays inline. Audio files in the file explorer (mp3, ogg, wav, and the rest) used to say "cannot be previewed". They now open with a real play button. Voice notes your employees generate play right there.
• Better document reading. When an employee cannot read a PDF or Office file directly, it now automatically falls back to extracting the text instead of giving up.

One quiet but important fix

Employees you imported or cloned from another instance could silently lose access to some of their own skills, because they carried over a stale security token from their old home. Importing and cloning now generate a fresh token automatically, so a cloned or imported employee works correctly from the first message.

Picking it up

Hard refresh (Ctrl+Shift+R) for the new version, and Chrome will update the browser extension on its own. A few of the employee-level changes (the plain-text questions, the new default skills) apply the next time an employee's session is cleared.

Want to test the most advanced AI employees? Try it here: https://Geta.Team