v2.4.9: Your Phone Agent Finishes Its Sentences, Opens How You Want, and Forgets by Default

Voice was the theme of this release. If you have your AI employees answering or making phone calls, v2.4.9 makes those calls steadier, gives you real control over how they open, and puts a hard boundary around what a phone agent can remember. There is also a macOS desktop refresh and one more white-label fix for good measure.

Here is everything that shipped.

Phone calls stop talking over themselves

If you ever put an agent on speakerphone, you may have watched it freeze mid-sentence, go quiet, or start answering its own words and wander off-topic. The cause was mechanical: the agent's own voice echoed back through the caller's microphone, and the system mistook that echo for the caller interrupting.

Calls now recognize and ignore this self-echo. When an interruption or transcript simply repeats what the agent just said, it gets discarded instead of cutting the agent off. On top of that, speech is delivered in more natural, complete units, so the whole call sounds smoother and less choppy. The practical result: your phone agent finishes its thoughts and holds its line, even in a noisy room.

You choose whether the agent speaks first

Greetings are now yours to shape, per phone number.

A new "Greeting (agent speaks first)" toggle decides whether the agent opens the call at all. Turn it off and the agent simply waits for the caller to speak. Turn it on and an optional text box lets you guide what the greeting should say, while the agent still phrases it naturally in its own voice. Leave that box empty and it uses the default greeting.

This works on both voice stacks (Gemini Live and ElevenLabs) and applies to inbound and outbound calls alike. Existing numbers keep behaving exactly as they do today, so nothing changes until you opt in when configuring a number.

One panel to configure a number's voice

Setting up a new phone number and editing an existing one now open the same rich configuration panel. Previously, first-time setup only exposed part of the options, which meant freshly configured numbers were quietly missing choices that only appeared later when you went back to edit them.

Now you pick the voice stack, the voice, the greeting, calendar access, Google Search, and custom instructions all in one place, whether you are assigning a brand-new number or adjusting one you already use. No more drift between two different forms that were supposed to do the same job.

Memory access is now a deliberate choice for phone agents

By default, phone agents no longer reach their memory database at all. A per-number "Allow memory access" toggle, off by default, controls it.

The important part is where the enforcement happens. When the toggle is off, access is blocked at the execution layer, not just hidden from the tool list. Any attempt to reach the memory database, through any general-purpose tool, is refused outright. That means a phone agent cannot recall information across calls unless you explicitly turn it on. For anyone with privacy or data-protection requirements, this is the difference between a setting you hope is respected and a boundary that actually holds.

Call recordings clean themselves off the provider

When call recording is enabled, each finished recording is now downloaded to your instance and then deleted from the telephony provider. The audio lives only in your own storage, paired with its call transcript, and never piles up in the provider's cloud.

This keeps ownership of your recordings where it belongs, on your infrastructure, and stops sensitive audio from quietly accumulating somewhere you do not control. It fits the self-hosted-by-default posture that runs through the whole platform.

macOS desktop app 1.0.10

The macOS desktop installer and its auto-update feed move to 1.0.10, seeded to every instance. Mac users are offered the new version automatically, no manual download required.

The license badge honors your brand color

Small but satisfying: the dashboard's employee-count and license badge used to stay blue no matter what theme you set. It now follows your white-label primary color, matching the focus-ring fix from the previous release, so it looks right in both light and dark mode.

On the horizon

A couple of things we are still working through for a future release:

• Voice loudness on ElevenLabs. We are continuing to investigate consistency on the ElevenLabs voice path. The telephony pipeline gives us no direct output-gain control, so the most promising lever is voice selection (choosing a consistently recorded voice) rather than a runtime setting.
• Post-call email quality. We are considering an option to have the higher-quality post-call summary email replace the in-call one.

The theme underneath

Most of v2.4.9 is about the phone, and phone calls are where an AI employee is most exposed. There is no interface to hide behind, no place to quietly retry, just a voice on a line that either holds up or does not. Making calls that do not cut themselves off, greetings you can direct, and memory boundaries that are genuinely enforced is what moves a phone agent from a demo to something you can actually hand a number.

If you are running Geta.Team, update your macOS app and take a fresh look at your phone number settings. The new voice configuration panel is where all of this comes together.