v4.0.9: Meetings Move to Push to Talk, Your Phone Agent Can Reach Its Whole Toolkit, and Admins Get Daily Limits

The headline change in v4.0.9 is one you will feel in the first meeting you run after updating: your AI employees stop answering when they hear their name.

Meetings move to push to talk

Saying an employee's name no longer makes it speak. That single change fixes two things at once. You can now talk about an employee in a meeting without it interrupting to help, and employees no longer end up answering each other in a loop while the humans watch.

The default is standby. An employee on standby listens to the whole meeting and keeps everything said as context, silently. When you want its answer, you hold the talk button on its tile. It catches up on everything said since its last turn, then responds.

Hold to talk and release when you are done, or click once to start and once to stop. Space still works on a computer. A short tip explains it the first time a meeting starts, so nobody has to be told twice.

If you prefer the old behaviour for a specific case, Conversation mode is unchanged: chosen people talk to an employee freely, no button required.

Your phone agent can now reach its whole toolkit

Until now, an employee on a phone call had the voice and little else. Ask it to look something up in its files, check a record, use a skill or prepare and send something, and it could not do it mid-call.

Now the voice can hand that request to the employee's chat session, which has every tool. It tells the caller it is checking, and the conversation carries on. When the chat session finishes, the answer is pushed into the call and the voice delivers it, waiting for the end of whatever sentence it is saying rather than cutting itself off. Past two minutes it says it is taking a little longer.

If the call ends before the answer is ready, nothing is lost. The chat session texts the answer to the caller when the caller was not you, and if it was you, it lands in your chat.

The chat session knows who asked, including whether it is the owner, and is instructed not to share anything private with anyone else. Because anyone can dial a public number, this is off by default and lives in each phone number's settings as "Allow requests to the employee during calls". In-app calls already worked this way.

Daily limits per user or per employee

Enterprise admins can now cap daily messages and daily call minutes, either for a user across all their employees or for a single employee. Both are set in Admin, on the user's panel under Users and on each employee under Manage Employees, which also show what has been used today. Empty means unlimited, which stays the default. Where a user and one of their employees both have a limit, whichever is reached first applies. Counters reset at midnight, server time.

Messages count when they reach the employee, whatever the source: a chat message, a scheduled task or heartbeat, an inbound email or channel message, a delegation from a colleague, or a task asked out loud during a call or meeting. Over the limit, the chat shows a clear notice with the reset time, automatic triggers are skipped rather than queued, and a task asked during a call is declined out loud.

Call minutes count time the employee spends actually talking: in-app voice and video, phone calls on Twilio and SIP in both directions, and meetings. A call will not start without minutes left. In the app you get a warning a minute before time runs out and then the call ends; a phone call is hung up; in a meeting the employee warns in the chat a minute ahead and then leaves, and the meeting carries on without it.

Users see their own limits and their employee's in the chat sidebar, the same way SaaS credits appear, updating the moment something is counted. Call time is tracked to the second, so even a thirty second call shows up.

This is Enterprise only. Free tier and SaaS have their own credit systems.

WhatsApp: the messages you send from your own phone

A new option in the WhatsApp assignment, off by default: "Also pass on the messages I send from my phone (private chats)".

Turn it on and when you write to a contact from your own phone, your employee receives the message for information, along with who you sent it to. An appointment you confirmed, a follow-up you promised. It will never answer that contact in your place. Messages the employee itself sent stay excluded, and it recognises its own even after a restart. Your contact allowlist applies exactly as it does for incoming messages, and group messages are unchanged.

Temporary accounts show their countdown

If an account has an end date, the user now sees it rather than discovering it at the door: in the sidebar, the dashboard and the chat, as "Access until..." with the time remaining. It turns amber at three days left and red on the final day.

In Admin, the date appears next to each user and is set with a date picker, or with quick +7 days, +30 days and "No expiry" buttons. The account works until the end of the chosen day. Accounts with no end date display nothing at all. When the date passes, sign-in is refused as before and the user's data is kept.

MCP servers tell you whether they actually work

A server could sit there installed and enabled while the employee's session quietly failed to connect to it. A wrong header name, a refused token, a wrong address. The tools simply were not there, with nothing anywhere to say why.

Every HTTP server now has a "Check connection" button, which also runs automatically right after you add one. It connects using the same address and headers the session uses, then reports either "Connected: N tools available" or the actual reason it failed: credentials refused, nothing at that address, no answer, not an MCP server.

Each server also shows how the employee's current session sees it, with the session's own error if there is one. A server you changed after the session started shows no session status until the next session, which is the honest answer rather than a stale one.

The memory bug behind a lot of small problems

Some instances with long conversations were running out of memory, and the symptoms did not look related: chats showing "Reconnecting..." all day even mid-task, WhatsApp dropping out so messages failed to send, and encryption keys drifting out of sync so group messages never arrived.

The cause was transcripts being read whole. At the end of every turn, for every employee, the server read all of an employee's transcripts end to end just to count tokens. The context gauge and the list of past conversations read them whole too. A transcript with inline images can reach hundreds of megabytes, so the main process would jump to around 2 GB within seconds, stall while it did, and on a smaller machine get killed by the system over and over.

Transcripts are now read streamed or from the end, never whole. The context gauge reads only the end of the conversation, the token count skips conversations not written in the period and decodes only the relevant lines, the conversation list reads only what was added since it last looked, and opening a past conversation no longer loads the file into memory twice. On a 300 MB conversation: roughly 2 GB at peak before, a few dozen megabytes now, with no stalling.

Also fixed

• Notifications that arrived while you were chatting took minutes to be handled. The wait after your employee's last answer drops from 90 seconds to 20, and each queued notification now starts as soon as the previous one finishes instead of waiting on a periodic check. A message you send your own employee on WhatsApp counts as a conversation, not a notification, so it no longer waits in line at all.
• Scheduled tasks kept firing on a cloned employee after being deleted. Clones and imports copied tasks with identical IDs, so one silently replaced the other. Tasks are now tracked per employee, and clones and imports get fresh IDs.
• WhatsApp could report a message as sent when WhatsApp had actually refused it. A send now waits briefly for a refusal and fails with the reason in plain words, and a later refusal marks the message failed and tells the employee. Delivery and read receipts now update message status properly.
• WhatsApp media failed for anyone an employee was shared with, because the access check used a column that does not exist. It now follows the same rule as the employee's other files.
• An outgoing call the employee tried to cancel kept ringing until the operator gave up. Ringing calls can now be cancelled on both SIP and Twilio, and the prepared voice session closes immediately.
• An admin with no employees of their own saw an empty employee picker in Connectors until they typed a search. Admins now see every employee on the instance straight away.
• Hebrew and Arabic on mobile had right-to-left glitches: the empty message field laid out its placeholder the wrong way round, the "..." menu opened past the screen edge, and several directional arrows pointed the wrong way. All three are fixed.
• On mobile, the dictation microphone was a 36 px button sitting next to the call button, so it was easy to miss and easy to start a call by accident. Microphone, call, send and stop are now 44 px touch targets, with the microphone moved further from the call button. Nothing changes on a computer.

Desktop App 1.0.58, Windows and Mac

F5 and Ctrl+R did nothing, so when an instance restarted while the app was open, the page stayed broken and the only reload hid in the tray menu. F5 and Ctrl+R (Cmd+R on Mac) now reload the active tab, Ctrl+F5 or Ctrl+Shift+R reload without the cache, and there is a refresh button at the end of the tab bar.

Two Windows white label fixes: the app no longer starts twice at login (older versions registered a second startup entry under the brand's identity, and the app now cleans that up once at next launch), and a pinned branded app keeps its own logo and name instead of showing the Geta.Team icon. An app pinned before this version needs unpinning and pinning again.

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