v4.0.25: Call Recordings Get an Expiry Date, the Mac App Records Calls Too, and the Cube Sets Itself Up
Call recording without a bot in the meeting launched last week on Windows only. The Mac can now do it too, from macOS 14.2 onward.
v4.0.25 closes two loose ends from the call recording launch, makes the Cube set itself up, and fixes a quiet bug that could move your Claude employees to a different model without anyone choosing it.
Call recordings now have an expiry date
When v4.0.24 shipped, the MP3 of every recorded call was kept beside its transcript, with no end date. Now the sound is deleted after a retention period, the same way meeting recordings already work: 7 days, 30 days (the default), 90 days or 1 year.
Instance admins set it in Settings > Admin Settings > Call recordings. Two things worth knowing:
- Only the audio goes. The transcript stays with the employee, along with everything the employee took from the call into its memory and your kanban.
- A change also applies to calls already recorded, immediately. Saving the setting recomputes the expiry of every recording still kept. Shorten the period and anything already past the new limit is deleted right away, so choose deliberately. Lengthen it and the recordings still there are kept longer (ones already deleted do not come back).
Each call in the transcripts view now shows how long its sound will be kept, then that it was deleted. For most teams, 30 days is about right: long enough to go back and check a figure someone said, short enough that you are not quietly building an archive of every conversation you have ever had.
The Mac app records calls too
Desktop app 1.0.81 brings call recording to macOS, the same as on Windows. Your microphone is you, and the other voices come from the Mac's own sound. That needs macOS 14.2 or later, and the app asks for the system audio permission the first time. The same rule as before applies: the app reminds you to tell people the call is being recorded, but telling them is your job.
The Mac app is now signed and notarized by Apple, and there are a few more Mac improvements in this version:
- The first-run setup walks you through each permission the app needs, including the restart macOS sometimes insists on.
- When the employee controls your computer, it can now read the elements on screen, so it clicks exactly where it should.
- The menu bar and Dock icons follow light and dark mode.
Windows gets the same version number, with the same features as before.
The Cube sets itself up
The Cube panel in Geta.Team now connects to a real Cube plugged into your computer by USB, in Chrome or Edge. There is no separate page and no new tab. It finds the model (240 or 480 px) by itself, stays connected while you move around Geta.Team, and reconnects on its own if you have already allowed it. The browser asks once which device to use, and the panel tells you which one to pick.
Setup is nearly hands-off. Choose the Cube's employee and type your Wi-Fi in the panel, then either:
- Plug it in. The Wi-Fi and the pairing go through the cable, and the pairing code the Cube shows is validated straight away. Nothing to type.
- Pair by sound. Unplugged, tap "Pair by sound" on the Cube and play the sound from your computer. The Cube hears your account's address, a one-use ticket valid for 10 minutes, and your Wi-Fi details, then joins the network and pairs by itself. The sound stops as soon as it is paired.
Who answers when you talk to the Cube
There is a new setting on the Cube for who handles your voice:
- Voice assistant (the default): the voice answers by itself using its own tools (memory, calendar, mail, files, the time, the Cube's screen) and hands real work, like a report, a deployment or an analysis, to the employee's chat.
- Employee: every request goes straight to the employee's chat.
Either way, what the employee writes is spoken as it arrives, first what it says before using a tool and then its answer, with nothing added. Next to the talk button, the Cube shows whose turn it is: yours, the employee thinking, or the employee speaking. A push to talk held for less than a second now sends nothing, so a stray tap no longer reaches your employee.
Telegram: several bots, one employee, and no more collateral damage
If you gave one employee two Telegram bots, say one per client, this release fixes three things that were broken:
- Replies went out through the wrong bot. Every reply used a single bot, so people writing to the other one never got an answer. Replies, text and media now go out through the bot the chat last wrote to (then a bot that knows the chat, then the most recently assigned one), and only if that bot is allowed to write to that chat.
- Removing one bot removed them all. "Remove" now removes only the bot you clicked, and changing one bot's permissions no longer changes the others.
telegram infoonly showed the first bot. It now lists every bot, each with its status and allowed chats.
Moving a bot to another employee also got easier. "Edit permissions" now has an employee picker: choose someone else, save, and the bot moves in one step. A bot answers for one employee only, so the previous employee loses it at the same moment.
The fix you might not have noticed you needed
Claude and Codex now keep the model you chose. Until now, the model picked in the chat only applied to the running session. After a restart or a resume, Claude fell back to the default of the connected Claude account, and the provider can change that default whenever it likes. In other words, an employee could end up on a different model without anybody choosing it.
The model you pick is now stored for the employee and used at every start and resume, for Claude and Codex. If you never chose one, Claude uses Opus (always the newest Opus), and Codex keeps its usual default. Fable is a Geta.Team setting: it is only used for Geta.Team administrators and for accounts where Fable has been enabled on Geta.Team. Everyone else runs on Opus by default, even if they pick Fable.
Exporting a big employee actually finishes now
Exports used to be built inside a single request. Past a few minutes, the browser or the proxy cut it off, nobody got the file, and a temporary copy as big as the employee stayed on the server for good.
Now the export runs in the background. "Export" answers straight away, the panel shows live progress, and your browser's own download starts when it is ready. You can close the panel and find the export again when you reopen it, or cancel it. The archive is written straight from the employee's folder with no full copy first, which halves the disk space it needs. Free space is checked before it starts, and a broken download resumes where it stopped. The file is deleted an hour after a complete download, or after 24 hours, and anything a crash left behind is cleaned up at startup.
Smaller fix
When a pending notification was waiting behind a running task, the "Process now" button stayed on screen but did nothing. It is now hidden while the task runs, with only "Waiting..." showing, and comes back, clickable and with its countdown, once the task ends. Same in the Studio panel.
Want to test the most advanced AI employees? Try it here: https://Geta.Team