v2.7.24: Community Edition Opens Up Gmail, Outlook and the Browser, Your Edits Reach a Running Employee, and Codex Takes Direction Mid-Task

This is one of the bigger releases in a while. Community Edition gets the connectors people reach for first. Editing an employee's instructions now actually reaches the conversation you're having. Codex employees get a batch of fixes that make them behave the way the picker says they will. A few quieter fixes cover fresh installs, desktop re-pairing and MFA.

Community Edition gets the browser extension, Google, Microsoft and Telegram

Community Edition was meant to be usable, and in one specific way it wasn't: the features people tried first were the ones that turned them away. The Chrome extension installed, paired, and then refused to do anything. Gmail and Outlook showed up in the interface and did nothing. The Connectors panel wouldn't even open.

Four things now work on Community Edition: the browser extension, Google (Gmail, Calendar, Drive), Microsoft (Outlook Mail, Calendar, OneDrive) and Telegram.

The surprising part is that nothing had to be added. We assumed Community Edition employees were being hired without the Google and Microsoft skills. They weren't. Those skills, and browser automation, had been installed on every employee all along. They were being blocked one layer higher up, by three separate locks sitting in three different places. From the outside that looked like one vague "Enterprise only" wall. All three are gone.

Slack, Teams, WhatsApp, Discord and phone calls stay Enterprise, and they're now refused properly rather than half-refused. In the Connectors panel those sections are greyed out with a clear lock instead of looking available and failing when you try them. Incoming webhooks from every provider stay open, so an Enterprise instance that gets downgraded doesn't start quietly dropping messages it has already been promised.

We also moved where the rule lives. The Enterprise restriction is now written in one place, five lines long, instead of being copied into each connector separately. Scattered copies are exactly how Google and Microsoft ended up locked without anyone noticing, so we'd rather not do that again.

Editing instructions now reaches the employee you're talking to

We owe you a straight correction here. In v2.7.23 we wrote that a running employee re-reads its instructions during a session, so saving them no longer needed a restart. We checked that on a live employee this week, and it isn't true. The instructions are read once, when the session starts, and never again on their own.

So after the last release, saving new instructions did nothing to a conversation already in progress. The employee carried on with what it started with, and the panel told you the change had applied. That's worse than the restart it replaced, because the restart at least lost your conversation where you could see it.

The test was simple. Ask a running employee for a code that doesn't exist, and it says it has none. Add the code to its instructions. Ask again in the same conversation. Before this release, it still had none.

Now, saving sends the employee a nudge: your instructions were edited, re-read them. It does, and it applies them straight away, with the conversation, the context and any open work intact. Run the same test through the real save button and the employee answers with the code you added seconds earlier.

The nudge only goes to employees that are running. One that's asleep picks up the new instructions when it next starts, since waking a container on every save would be wasteful. If the nudge fails to send, the save still goes through.

Codex employees: the right model, steerable mid-task, and a working context meter

A group of fixes so Codex employees do what the interface says they're doing.

• New conversations start on the model the picker shows. The picker displayed one model while Codex quietly used its own default for the first message. Switching models mid-conversation already worked; starting correctly didn't. Now it does.
• The model list follows your account. It used to be a fixed list inside Geta.Team, so it drifted from what your Codex account actually offers. It now comes from your account, with reasoning levels that match the model you pick. (Claude's Fable label, which still said version 5, now correctly says 5.1.)
• You can steer a task that's already running. A message sent while Codex was working used to wait in a queue and start a separate turn afterwards. It now goes into the task in progress, so "actually, use the March figures" lands while it still matters. If Codex finishes in that same instant, the message falls back to the queue rather than getting lost.
• The context meter works for Codex. It was looking for usage in the wrong folder, so the percentage and the "compact this conversation" suggestion never appeared. Both show up now. The suggestion appears at 50% and turns critical at 85%, and /compact runs Codex's real compaction, holding any messages that arrive until it finishes.
• Stop interrupts the task again. Two releases ago we said stop means stop. On Codex employees, the current version needs both the conversation and the specific task identified before it will interrupt, and we were only sending the first, so the interface went idle while the employee kept working. Now both are sent. A stop pressed in the split second before a task has its ID is held and applied the moment it does. In a live test, a 30-second task stopped in about two seconds, with nothing left running.
• Tool calls survive a page refresh. Newer Codex conversations store tool calls in a different format, and the replay ignored them, so they vanished on refresh. Both formats are read now, and images inside tool output are left out of the replay rather than flooding your screen.

One small fix applies to every conversation: a very long command in a tool row could make the whole chat wider than its panel and add a horizontal scrollbar. Long rows now stay inside the panel and truncate neatly, down to phone width.

A fresh instance asks for a model before it asks you to hire

A brand-new instance used to send its owner straight into the Employee Catalog even when no language model was connected. You could go through hiring and end up with an employee that couldn't work, while the setting you actually needed sat hidden behind it.

Now, until Claude, Codex or a custom model is connected, hiring is greyed out and Settings opens with direct shortcuts to API Keys and LLM Providers. The moment a provider connects, everything unlocks, without a reload. Gemini is no longer required at startup either. It's still there for the features that use it, but any of the three is enough to start hiring.

Desktop re-pairing keeps OS control switched on

If you revoked Geta.Team Desktop and paired it again without restarting the app, it looked connected but refused every OS action with "not armed". Bringing the window to the front didn't help, because focus has nothing to do with it.

Pairing now switches OS control on, including when the app reuses its existing connection. The "Allow OS control" switch in the tray and in Settings is still the master switch, and the error message now points to it. The desktop now also reports whether OS control is on, so the status you see is the real one. This ships in Windows version 1.0.13. The macOS build wasn't part of this release.

MFA can no longer be switched on in a way that locks you out

Two MFA settings could leave an instance either less secure than you thought or impossible to log into.

• Email codes are now enabled only after a real test code reaches you and you confirm it. If email delivery fails at login, you get a clear message saying the code couldn't be sent, instead of a generic login failure.
• Authenticator apps now require you to scan the QR code and enter a valid code before the mode becomes active, instead of deferring setup to your next login. Setup sessions expire after ten minutes and can be used only once.

Changing the MFA mode also resets trusted devices, so the new policy applies to everyone straight away.

What we took from this one

Two items in this release are the same bug in different clothes: the product said one thing and did another. The code comments said the connectors were Enterprise-only while the real lock sat somewhere else. The panel said your instructions had applied while the employee was still reading the old ones.

Both fixes came from the same two habits. Write a rule down in one place, where you can see all of it. And test a claim on a live employee before telling anyone it's true, including when the one telling is us, on this blog.

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