§0Rolling out now · desktop + mobile · announced Sept 17, 2026

Grok Bot Voice Notes: The Full Guide

Grok Bot voice notes are rolling out right now, and this is the full guide to what they do and how to get the most out of them.

Imagine your agents telling you what they finished while you’re out on a walk, and you approving the next move with one word.

Stick with me, because the biggest win has nothing to do with the voice settings.

The whole idea in one shot: your agents send voice notes, you approve by voice while you’re away from the desk, and the work moves.

§1Here’s what happened

September 17th: Grok Bot can talk

Five weeks from launch to a voiceAug 11beta launchAug 26plans widenSep 3enterpriseSep 16Voice on falSep 17Bot can talk
Tweet 1 · the launch

The official demo

This is the Grok Bot account’s own post: five words and a short demo video. Watch the panel in it — a live waveform, a mic control and a hang-up button — because the rest of this guide keeps coming back to it.

Tweet 2 · the rollout

Desktop and mobile, over a couple of days

The follow-up post says voice is rolling out on desktop and mobile over the next couple of days. It is staged, so if it hasn’t landed on your device yet, that’s expected.

Tweet 3 · the confirmation

Elon Musk, a few hours later

Musk quoted the demo a few hours after the launch post with one line: Grok Bot now has a voice. That’s the third confirmed piece, alongside the demo and the rollout window.

Straight from the sources — open everything ↓
§2How it works

Two things are happening at once

Two things are happening at once1 · you speak to your bot — a live call2 · your bot speaks to you — a voice noteYoutalk · listenYour botGrok Voice
§3Thing one · you speaking to your bot

A call-style panel

It behaves like a phone calllive waveformmicrophonehang up

This is a diagram of what the official demo shows, not the app itself — the real panel is in the launch tweet above.

§4Thing two · your bot speaking to you

That’s the voice note

Nobody has to be sat there at the same timeBot recordsa shortaudio messageSends it overit waits for youYou listenwheneveryou want
§5What powers it

Grok Voice Think Fast 2.0

Grok Voice Think Fast 2.0 — xAI's reported numbersTime to first audio · previous Think Fast1.25sTime to first audio · Think Fast 2.00.70sSpeech accuracy vs the previous version1.4×Accuracy vs Deepgram Nova 3 · Scribe v21.5–2×
numbers as reported by xAI · announced July 29, 2026 · source
Tweet 4 · the engine

The Grok account, an hour after launch

Grok’s own account replied to the launch with four words: powered by Grok Voice. Both halves — the call and the voice note — run on the same voice model.

§6Number one · speed

Speed decides whether it feels natural

Speed decides whether you keep using it0.7 secondsyou forget it’s software1 second — the linestill feels natural3 secondsyou talk over it
§7Number two · accuracy

Accuracy decides whether it’s safe

When an agent can act, a misheard word costs moreOne misheard wordAgent only writes backa clumsy sentenceAgent acts in your appssomething real changes
THINKING IT? “What if it mishears me and changes something it shouldn’t?”

That’s exactly what the approval prompt in the next section is for.

Anything that can’t be easily undone should ask you first, and with voice, saying yes takes one second.

§8The most important detail in this release

The bot asks for approval

In the demo: a board deck reviewBot plansschedule thedeck reviewBot asks youby voiceYou say yesone wordIt actsthe calendarchanges
§9Ryan Merket on the stakes

Same voice. Completely different consequences.

Two spoken requests, two very different endingsYou speak a request“Summarise this doc”ends with an answer“Update the account”a real action,outside the chat
Summarising a document ends with an answer. Updating a customer account, sending an email or changing a calendar produces a real action outside the conversation.paraphrasing Ryan Merket · RuntimeWire, Sept 17, 2026
§10Why the approval matters

Fewer agents sitting frozen, waiting on you

The approval prompt is the control layerYou — answering by voice, instantlyApproval prompt — asks before it actsSoftware you’re already logged intoCRM · email · calendar
§11The piece underneath all of this

Its own persistent computer in the cloud

Every account gets its own computer in the cloudYour bot’s cloud computer — always onBrowserlogged inFilesystemkeeps filesTerminalruns coderole context · repeatable workflows · scheduled routinesYour laptopclosedit keeps working
real Grok Bot app · recorded on my machine, Aug 2026

What you’re watching: the panel on the right is that bot’s own screen on its cloud computer, with its scheduled routines listed underneath.

§12Already wired in

Teams, Salesforce, HubSpot and Gong

Where the bot can already search and actMicrosoft Teamssearch + actionsSalesforceHubSpotGongYour botworks insideyour apps
sales researchsupport queuesrecruitingexpense managementbug reproduction
real Grok Bot app · recorded on my machine, Aug 2026

What you’re watching: the real plugin marketplace inside the app, sorted by category, with an Add button on every one.

It’s not reading you a summary. It’s telling you what it did while you were away from your desk.
§13Getting value out of it

Three things separate real use from two days of playing

The three things1 · Context inbrief it likea handover2 · Notes backshort, becauseit knows you3 · Voicesa filing systemfor your ears
§14Thing one · context on the way in

Talking is about 3× faster than typing

Talking is around three times faster than typingTyping a briefSpeaking a brief≈3×
§15Why typing holds you back

Short briefs produce vague output

The loop typing traps you inTyping onlyphone keyboardShort briefone lineVague output“not very good”Back to DIYyou do thejob yourself
§16The same brief, two ways

Type it and you write one line. Speak it and you say more.

OLD WAY · TYPED
1 line
research competitors for the acme pitch
  • No idea who the customer is
  • No idea what you already tried
  • No idea what you don’t want
  • Vague output comes back
  • You decide the agent isn’t very good
NEW WAY · SPOKEN
≈3× the detail
“Look into competitors for the Acme pitch. They’re a 40-person logistics firm and they care about price. I already checked the two big names, so skip those. And don’t give me a list of features — I want where we win.”
  • Who the customer is
  • What you already tried
  • What you don’t want
  • Usable output comes back
  • Same minute of your time

The Acme brief is an illustration of the pattern, not a real client.

§17Treat it like a handover, not a command

Say what a bad answer would look like

Three things to include every time you speak a briefThe backgroundsay it like theysit next to youWhat worries youthe thing you’dnever typeWhat bad looks likethe fastestquality upgrade
§18Thing two · what comes back

Where most people will go wrong

How the feature quietly leaves your dayThin contextit barelyknows youWaffly updatere-explains whatyou knowYou skip onethen the nextHabit goneyou stoplistening
§19What good sounds like

Thirty seconds, not three minutes

A good voice note is shortThin context — it has to set the scene3 minRich context — straight to the point30 secWhat it didWhat it foundWhat it needs
§20In my own setup

Every agent reads and writes to one shared vault

Nothing gets re-explainedResearch agentwrites findingsContent agentalready knowsOutreach agentreads the sameReport to meshortOne vaultshared memory
live dashboard · my Agent OS

What you’re watching: the memory layer inside my own Agent OS — one vault every agent pulls from, so reports back to me start from what’s already known.

§21The standard to hold Grok Bot to

A clear role, the right tools, a record of how you like it

What makes its voice notes tightClear roleone jobRight toolsaccess it needsYour wayshow you likethings doneTight notesyou keeplistening
real Grok Bot app · recorded on my machine, Aug 2026

What you’re watching: a real bot being created from nothing — the plus button, create new bot, and the full brain dump of its role going in.

§22Thing three · voices as a filing system

One bot, it’s decoration. Several, it’s how you cope.

real Grok Bot app · recorded on my machine, Aug 2026

What you’re watching: the real Grok Bot app on my machine. Every row in the sidebar is a different bot with its own job, and now each one gets its own voice.

§23Kevin Niparko’s live demo

A five-bot product team

Five bots. Five voices. Know who’s talking in two words.CoraEmilyLunaPetePixelthe roster Kevin Niparko demoed live at the Grok Bot event
real Grok Bot app · recorded on my machine, Aug 2026

What you’re watching: two real bots in one group chat handing work to each other. At five bots, the hard part is knowing which one is talking to you.

§24How to choose

Not the nicest voice. The most different one.

Pick the voice that sounds least like the others✕ picked because they sound nice — you still check names✓ picked to sound least alike — you know within two words
§25Quick one before we keep going

Get agents that report back to you — across your own tools

live dashboard · my Agent OS

What you’re watching: the Agent OS dashboard — every agent in one place, sharing one memory layer.

Over 3,000 business owners · plenty had never used AI before they joined

Get more leads and customers — with agents that hand work between themselves.

There’s an agent operating system in there you download as a zip file. Plug in whatever you’re running — Grok Bot, Claude, Hermes, OpenClaw — into one dashboard with a shared memory layer, so your agents stop repeating each other.

The Agent OS as a zip file — one dashboard for Grok Bot, Claude, Hermes and OpenClaw
A shared memory layer — so your agents stop repeating each other and start handing work over
A 30-day roadmap — to get it live, plus a full video walkthrough
Daily updates — as we improve it
Four coaching calls every week — bring your own agent setup and we fix it live
Which jobs are safe to hand over — and which ones need an approval step in front of them
Join the AI Profit Boardroom → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description, or search the AI Profit Boardroom on Skool
THINKING IT? “Doesn’t running an Agent OS burn a fortune in tokens?”

No. The everyday work runs on free local models and free APIs, and for the heavy work it drives the CLIs you already pay for, like the Claude CLI that comes with your Claude subscription.

Inside the Boardroom there are full token-efficiency tutorials too, so you learn to cut usage right down.

§26What this changes in a normal working day

Listen while you’re walking, driving or working

Listen while you’re walking, driving or workingThe walkfour updatesThe drivehands on the wheelThe gymeyes elsewhere
Instead of stopping what he’s doing to read routine updates, he can listen while he’s walking or driving or working.paraphrasing Michael Fenech, who’s been using these daily
§27Why it’s bigger than it looks

Most agent setups fail at the review

Agents don’t fail at the work. They fail at the review.Research · doneContent · doneOutreach · doneSupport · doneYour screen timeone windowat a timeDecisionsapprove · fixor bin
§28Audio moves the decisions

The walk. The gym. The drive.

By the time you sit down, the work has already moved4 updatesplayed onthe walk2 approvedby voice1 redirected“do it thisway instead”Work movedbefore yousit down
§29More decisions per day

The limit is attention, not computing power

You can only look at one window. You can listen to a run.Eyes — open, read, close, nextEars — four short updates, back to back
§30The limits

What SpaceXAI hasn’t published yet

What’s confirmed — and what hasn’t been statedCONFIRMEDNOT PUBLISHED YET✓ The demo✓ Desktop + mobile rollout✓ Musk’s confirmationWhich languages ship first?Every workflow on day one?Pricing for voice sessions?

Don’t build a plan around anything that hasn’t been stated.

§31A related release · September 16th

Grok Voice on the fal developer platform

Shipped inside the product and underneath itInside the product — Grok Bot voiceSept 17Underneath — Grok Voice on falSept 16 · developers build their own voice agentsThe same underlying Grok Voice model
fal · Grok Voice speech-to-speech API · launch coverage
§32A useful signal

Voice quality won’t be what separates setups

What separates a useful setup is what sits under the voiceThe voice — getting easy to getMemorywhat it knowsAccessthe right toolsClear rolesone job eachApprovalin the right places
§33The filter before any real task

Easy to spot, easy to contain, easy to undo

Three questions before you hand anything overSpot it?can you tell ifit went wrong?Stop it?can you stop itpartway?Undo it?can youreverse it?3 yeses → let it run1 no → keep the approval prompt
§34Putting the filter to work

Research is three yeses

real Grok Bot app · recorded on my machine, Aug 2026

What you’re watching: a real research bot being briefed — public web only, never touch my email, calendar or files. If research goes wrong you’ll see it, and nothing broke.

Three yeses — hand it over

competitor informationdrafting outreacha summary before a call

One no — it asks before it acts

writes into a live customer recordsends something on your behalfmoves a calendar others rely on
§35Where this goes

The next step might be hardware

Once your agent can speak and listen, the screen is the frictionToday’s frictionphone out,open the appVoicespeak + listenNo screen at all?Fenech’s read:hardware next

We’re not there. But you can see the direction from here.

§36Let me pull all of this together

Grok Bot voice, on one screen

i.
Rolling out nowDesktop and mobile, staged.
ii.
Think Fast 2.0Faster to first audio, 1.5–2× the speech accuracy of rivals.
iii.
A call-style panelSpeak assignments like a phone call.
iv.
Voice notes backListen away from a screen.
v.
A voice per botA roster you can manage by ear.
vi.
It asks permissionBefore it changes a real system.
§37To get the most out of it

Four moves

1
Brief it like you’d brief a person — not like you’d type a command.
2
Give your agents enough context — so their updates can be short.
3
Give each bot a voice that sounds nothing like the others.
4
Keep an approval step — in front of anything that can’t be easily undone.
still unknown: languagesstill unknown: voice-session pricingstill unknown: every workflow on day one?
§38Three beliefs to drop

What’s holding people back

Wrong: The best voice settings will make this useful.
Right: Memory, access, clear roles and approval steps make it useful. The voice is the easy part.
Wrong: Voice is just a faster way to give orders.
Right: It’s a handover going in, and your agent asking permission coming back.
Wrong: I can’t supervise more agents without more hours at a screen.
Right: The limit is attention, and your ears are free on the walk, at the gym and on the drive.
Don’t take my word for it

Members post their wins every day — agency owners, ecom founders, course creators and solo operators across 38 countries.

Read the 158-page wins doc →
§39If you wouldn’t know where to start

That’s the exact gap we close

Over 3,000 business owners · most of them starting from nothing

Get more leads and customers — with updates that are short and worth listening to.

You get the agent operating system as a zip file, so you can run Grok Bot alongside Claude, Hermes and OpenClaw in one dashboard, all sharing one memory layer.

The Agent OS as a zip file — Grok Bot, Claude, Hermes and OpenClaw in one dashboard, one memory layer
A 30-day roadmap — for getting your first agents live and reporting to you
Daily tutorials — agents for research, content, outreach and AI avatars — you get all of it
Four coaching calls every week — play back what your agents are telling you and get it tightened up live
A prompt library — for writing briefs your agents don’t misread
A member map — find people near you doing the same thing, with someone always online
Join the AI Profit Boardroom → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description, or search the AI Profit Boardroom on Skool
The people who get the most out of this won’t have the best voice settings. Their agents already had something worth saying.