Grok Bot vs Muse vs Dots vs Claude Duet: Four Always-On AI Agents Compared

Last verified: 24 October 2026, 14:10 UTC. Grok Bot and Muse both changed their plans or regions within weeks of launch, and Duet is still a research preview. Check the vendor's page before you pay.

By Frankie C. · Senior Market Researcher, Mindber. Tracks 500+ AI/SaaS tools across SEA markets via the Mindber Innovation Index methodology.

How we assessed this: We read each vendor's launch post, help pages and pricing page, plus Grok Bot's documentation and changelog. We also read the reporting on each agent's first problems and a 532-comment Hacker News thread on Duet's launch. Posts on X by people who run more than one of these agents filled in the rest. The screenshots were captured at 14:10 UTC on the day of our check, and the yellow highlights are ours. Mindber has no seat in Duet's waitlisted preview and ran none of the four agents, so nothing here is hands-on testing. We haven't scored them on the Mindber Innovation Index or the Mindber Functionality Score either. Grok Bot ships a new beta build most weeks and Duet opened three days ago, so a score would be stale within days. Brand logos come from the open-source Lobe Icons set and belong to their companies. Muse and dots appear with the Meta and OpenAI marks. The cover banner is an AI-generated illustration, not a screenshot.

Grok Bot, Meta's Muse, OpenAI's dots and Anthropic's Claude Duet are the same kind of product from four companies. Each is an always-on AI agent with a cloud computer of its own. It keeps working after you close the app, messages you when something needs a decision, and asks before it sends a message or spends your money. All four launched between 11 August and 21 October 2026.

Our pick for work is Claude Duet, if you live in the US and already pay for Claude. It costs nothing on top of a Pro plan, and its 6,000-plus connectors are the largest catalog any of the four publishes. Its approval rule fits in one line: nothing gets sent, paid, signed or deleted without your yes. Outside North America neither Duet nor Muse is on offer, which leaves dots and Grok Bot in most countries, Southeast Asia included. Someone whose work already runs through ChatGPT, Slack or Teams fits dots best, because a dot carries its context across all three. It starts with the $100 Pro plan. Grok Bot costs $30 and lets one person run a small team of bots. Every bot on an account shares one computer, which makes it the wrong place for a bank login. Muse is free for households in the US and Canada, and it has had the roughest start of the four.

The four agents side by side

From each vendor's own pages. Prices are in the chart in the next section.

	Grok Bot	Muse	dots	Claude Duet
Company	xAI, part of SpaceXAI	Meta	OpenAI	Anthropic
Status	Beta	Launched, with paid plans in limited testing	Rolling out to eligible plans	Research preview, invitation by waitlist
Model	Not named on xAI's Grok Bot pages	Muse Spark, with Muse Wonder coming to subscribers first	GPT-6 Astra	Claude Opus 5.5 only
Where you talk to it	Desktop app for Mac, Windows and Linux, iPhone, iPad and Android, Slack for Team Bots, or a tag on X	Muse app for iOS and Android, muse.ai and WhatsApp	ChatGPT on desktop, web and mobile, Slack and Teams, plus texting in a small US beta	Duet app for iOS and Android, and claude.ai
Connected apps	Marketplace plugins, plus its own browser for sites without one	Meta's apps, a small-business set and custom connectors	4,000+ apps through ChatGPT plugins	6,000+ connectors and plugins from the Claude Marketplace
Plans that include it	X Subscription Premium, Super and Ultra, Teams seats, and older Cursor and SuperGrok plans	Free tier, Power and Maximum	ChatGPT Pro and Business Premium, plus an Enterprise beta that admins switch on	Claude Pro and Max, not Team or Enterprise yet
Where you can get it	xAI's pages list no country limits	US and Canada	Pro: every market except the EEA, Switzerland and the UK	US only

Read the last row first. It rules out two of the four for anyone outside North America. A waitlist spot won't change that, because Anthropic ties Duet access to the email on a US Claude account. The model row tells you less than it seems. Duet runs only on Opus 5.5 and dots on GPT-6 Astra, while xAI doesn't say what powers Grok Bot. We couldn't find it named anywhere else either.

Function by function

From each vendor's docs and help pages. A dash means the pages we read don't say, not that the function is missing.

YesYesPartlyPartlyNoNoNot statedNot stated

Function	Grok Bot	Muse	dots	Claude Duet
Works on its own
Keeps working with the app closed	YesCloud computer, 24/7	YesRuns on its own VM	YesOwn cloud computer	YesOwn cloud VM
Messages you first	YesPrimary Bot offers work	YesUnprompted suggestions	YesProactive research	YesMorning brief
Scheduled and repeat tasks	YesRoutines	YesScheduled tasks	YesRecurring checks	PartlyYou set check-in topics
Learns a job from one demo	YesFollow-along routines	Not stated	Not stated	Not stated
Several agents at once	YesBots in group chats	NoOne Muse	PartlyOne dot, more planned	NoOne agent, many channels
See and steer
Live view of its computer	YesWatch and take over	PartlyActivity log, system files	YesOpen its computer	NoNot in the preview
Can use your own computer	YesOpt-in, asks each time	PartlyPaired Mac app (reported)	YesOpt-in, off by default	NoCloud only
Approval levels you choose	YesAsk first or allow rules	YesOnce, task, site, always	YesFour Custom Rule levels	PartlyOne fixed rule
Steps it always hands back	YesPasswords, 2FA, payments	Not stated	YesPassword changes, transfers	YesNo trading or transfers
Talk and share
Voice	YesLive voice chat	Not stated	YesVoice calls	YesVoice messages
Shared with a team	YesTeam Bots	NoPersonal plans	YesBusiness Premium	NoNo Team plan yet
An inbox of its own	YesBot email address	Not stated	NoNot at launch	Not statedSends from your accounts
Money, logins and memory
Pays for things	YesX Money or Link	YesLink, buyer protection	PartlyCustom Rules cover purchases	YesLink or Shop Pay
1Password sign-in	YesRolling out	YesRolling out	PartlyComing soon	PartlyComing soon
Forget one specific thing	Not stated	YesTell it to forget	NoDelete the whole dot	YesEdit Claude memory
TallyYes = 1, Partly = ½	14/15	8/15	10.5/15	7.5/15

Source: docs.x.ai, about.fb.com, Meta Help Center, openai.com, OpenAI Help Center, claude.com and Claude Help Center, Oct 25, 2026

Grok Bot documents all but one of the 15, and Duet the fewest at 7.5. Read the tally as reach, not as a ranking. Much of Grok Bot's lead rests on the one computer its bots share. Duet is still a brand-new preview, with a core Anthropic says it built around work first. Muse's four dashes are gaps in what Meta documents, which is a reason to test it on low-risk jobs before you hand it your inbox.

Each maker's own demo

All four launch videos run under three minutes, and each is the vendor's pitch. Watch them to see what the apps look like, not as proof that the agents do the job well.

1:47

2:37

2:28

1:39

Who can sign up, and what it costs

Where you live narrows the list before price does. In the US, all four are open to you, waitlist permitting. Canada gets Muse and dots. Southeast Asia gets dots alone. An individual in the EU or the UK gets none of those three, because OpenAI keeps dots off Pro plans there. Grok Bot fits into every one of those lists as far as we can tell, since xAI's pages set no country limits.

The sticker price hides four different ways of counting work. Duet shares the five-hour and weekly limits of your Claude plan, and its morning brief and background checks count against them. Anthropic says that is why Duet checks your accounts every few hours instead of constantly. Everyone admitted to the preview gets one full limit reset, usable until 30 November 2026. OpenAI went the other way with dots: talking to your dot doesn't count toward ChatGPT limits, while tasks it starts in Codex or ChatGPT Work do.

Meta meters Muse in tokens. Its subscription page gives Power 500 million Muse tokens a week and Maximum 3 billion, and says nothing about the size of the free tier. Grok Bot draws on one weekly pool shared with the Grok apps and Cursor, so a heavy week of coding in Cursor leaves less for your bots. Past the pool, extra work is paid from prepaid credit.

Teams pay differently again. A ChatGPT Business Premium seat costs $125 a month, Grok's Teams seats start at the Premium price, and Duet has no team plan yet. ChatGPT Plus, the plan most people weigh against Claude Pro in our Claude Pro vs ChatGPT Plus comparison, doesn't include dots at all.

What people using several of them report

We searched for a test that runs the same jobs through all four agents and found none. What exists are accounts from people who use several of them, and those are worth reading as experience, not proof.

Peter Yang, who makes AI tutorials for close to 300,000 followers on X, posted a review of all four on the night Duet launched. Duet comes out as the most proactive out of the box, with what Yang calls arguably the best model behind it, though the preview can't drive its own browser yet. Muse gets the nod for the best consumer experience and for business calling, but Yang rates its Muse Wonder model below Opus and Astra. Grok Bot's team of bots is the easiest to understand, in Yang's view, and its free access to X helps with research. Its email feature draws the one complaint. As for dots, it works inside ChatGPT and can drive all your threads there, yet Yang still finds it buggy and slow to start a voice call. The post also sets a limit worth copying: "I don't want more than 4 at a time connected to my email."

The embed cuts the post after Muse. The full post rates all four and links Yang's Duet video. Time shown in UTC. Source: x.com/petergyang/status/2113042717270049028

Yang's review video, linked from that post, walks through Duet and how it stacks up against the other three.

17:25

Two hours earlier, 宝玉 (@dotey), an AI engineer with about 260,000 followers, had posted a longer comparison in Chinese. That post's split of who each agent suits matches how the vendors pitch them. Muse is for everyday errands in the US. Duet is for owners who run their business on Gmail, Slack and accounting software. OpenAI's agent suits companies already on ChatGPT and Teams, and Grok Bot anyone who wants several AIs splitting the work. It quotes a reply that asks what the difference between them even is. Fair question, when all four launch posts promise the same thing.

The highlight reads: four companies have all launched the same kind of product. The embed cuts the rest of the comparison. Time shown in UTC. Source: x.com/dotey/status/2113014531853365681

The Hacker News thread under Duet's launch read like a small user survey. One commenter credits dots with handling almost all of a car purchase and is now using it to buy a house in Kyoto. Another finds Muse nagging about emails that don't matter, says it keeps getting blocked from Amazon, and wants to try dots once it is cheaper. A third, minimaxir, points out that Duet skips the pet-like characters of Muse and dots, which makes its use cases read as more professional.

Where each agent works, and how much of it you can see

All four keep a computer in the cloud, so the work goes on while your phone is locked. They differ in who shares that computer and whether you can look inside it.

Grok Bot

One persistent computer per account

• Every bot on the account works on the same cloud computer
• Each bot gets its own screen there, not its own security boundary
• You can watch the shared desktop live and take control
• Work on your own Mac or PC needs your permission, and the default is to ask every time

Muse

Muse Secure VM, one per person

• A dedicated cloud machine holds the agent, your data and your connected logins
• The agent runs in an isolated container, with Meta's safety services outside it
• Meta promises a Confidential VM, encrypted with a key only you hold, later this year
• Work continues after you close the app

dots

Its own cloud computer and browser

• You can open the dot's computer at any time and work in it
• Your laptop is optional, starts switched off and can be revoked
• It can start Codex cloud tasks in environments you set up
• Activity View shows its background work as it happens

Claude Duet

A cloud virtual machine with its own browser

• Starts with the memory, skills and connectors on your Claude account
• You can't watch its browser or take it over during the preview
• It can't see files on your own computer, and it can read but not change your Claude work
• Finished files land in a Make tab, and only you can mark a task done

Grok Bot's shared computer is deliberate, and xAI says so plainly. Every bot you run can use the browser cookies, files and command-line credentials stored there. The documentation tells you not to put a credential on it that another bot shouldn't use. The upside is handoffs: a research bot's downloads are waiting when the writing bot starts. The cost is that a login you give one bot is a login all of them hold. Those computers run in Cursor's cloud. SpaceX, which owns xAI, agreed in May to buy Cursor, as our SpaceX IPO analysis explains.

xAI's own description of the shared computer, including the line that the screens are work surfaces, not security boundaries. Source: docs.x.ai/grok-bot/computer-and-apps

Muse and Duet sit at opposite ends on openness. A Hacker News user with a seat in Duet's preview asked it to list the files on its virtual machine, and it refused. Muse, the commenter wrote, has a button in its phone app that shows its system files, and it will help you dig through them.

A comment from the same thread, posted the day Duet opened. Source: news.ycombinator.com/item?id=50188834

Openness has a price as well. In September a developer, Peter James, asked Muse to archive the files it could see and send them to Google Drive. James got 6.8 GB unpacked: what appeared to be the root filesystem of the session, Muse's internal documentation, memory files, agent logs and SSH key files. Meta's bug bounty marked the report Not Applicable, and James says it was never established whether the keys worked.

What each one does without asking

None of the four will send an email or spend money on its own by default. They differ in how far you can loosen that, and in who gets to decide.

Duet's rule is written into Anthropic's launch post. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you the exact action and waits. Some actions stay blocked even if you approve them: trading securities, moving money between accounts and getting into someone else's account. Anthropic also added a classifier that refuses to send credit card or government ID numbers. Health and financial details go out only when you have told Duet to send them.

OpenAI gives you more dials. A dot starts with built-in rules, and Custom Rules let you set each kind of action to run without asking, run if pre-approved, ask first, or hand off to you. A separate check called auto-review screens anything that could touch your accounts or share information. Changing a password or transferring money always comes back to you to finish by hand. While a dot researches on its own initiative, its tools can only read.

Meta built a second agent into Muse. Sentinel runs on the same machine, and per Meta's security write-up nothing Muse does reaches the internet unless Sentinel allows it. Approval requests arrive as a dialog in the app, outside your conversation with Muse. A grant can last once, for the session, for the task, for a set time or forever, and for email you choose whether Muse may only read or also send.

Grok Bot leans on rules you write. Its Auto Review settings take Ask first and Allow automatically rules, and Ask first wins when both match. Approval cards offer Always allow, which saves a matching rule. When the request comes from work that started without you, such as a routine, the card expires after about ten minutes and the action doesn't run. Since 23 October, an email or Slack message you ask for in your own words goes out once Auto Review clears it, while sends from routines still wait as drafts.

Passwords get similar care everywhere. A dot uses saved passwords without exposing them to the model. With Duet you type a password into a separate form on your phone, which sends it straight to Duet's browser and keeps it out of the chat. Muse swaps stand-in tokens for the real credentials at the network edge. Grok Bot hands you the screen for passwords and two-factor codes, and its new Vault can store logins and cards. For checkout, Duet and Muse both pay through Link, Stripe's wallet, which issues a one-time card number. Meta says Muse is the first agent covered by Link's purchase protections. Duet adds Shopify's Shop Pay, and Grok Bot can pay with X Money or Link.

1Password opened its agent sign-in on 22 October, with support for Muse and Grok Bot rolling out first. You approve each login in 1Password and it fills the password in, so the agent never sees it. Duet and dots are listed as coming soon.

Android Central's summary of the 1Password launch. Time shown in UTC. Source: x.com/androidcentral/status/2113309455127564451

What each agent remembers, and how to make it forget

All four learn from what you tell them and from the apps you connect. They part ways on undoing it.

A dot shares memory with ChatGPT in both directions and forms memories of its own from connected apps, even about things you never asked it. OpenAI's privacy FAQ is blunt about the limits. You can't view, edit or delete individual dot memories, and disconnecting an app doesn't remove what the dot already learned from it. Deleting the dot is the only reset. Turning off Memory in ChatGPT stops the sharing from then on, without erasing what the dot already has.

Duet begins with the memory, skills and connectors on your Claude account and can search your past Claude chats. Claude's memory help page says Duet also adds to that memory, and you can delete single memories or tell Claude what to forget. You can switch that search off in Settings under Search and reference chats. Disconnecting an app stops new access, but what Duet learned from it earlier stays.

Muse takes a forget command too: Meta says you can tell it to forget specific things it has learned. An earlier version of this page said Muse was the only one of the four with a documented forget. Claude's memory page shows the same control for Duet, so we corrected it. Meta also says your conversations and the data on your VM never reach its ad systems.

Grok Bot keeps memory per bot, covering preferences, role context and summaries of past work. Deleting a bot removes its profile, conversation and routines, yet xAI's FAQ warns that files and logins on the shared computer may remain. Backend retention follows Cursor's terms.

What has gone wrong so far

Three of the four drew public complaints or worrying test results within weeks of launch. Duet has been out for three days, so its record is still blank. Each case below says how solid the evidence is.

• Muse and Messages (reported, disputed). Jason Aten, a columnist at Inc., says Muse synced more than 187,000 rows of Messages history from Aten's Mac. Aten had declined that access during setup, and Muse then claimed it only saw notification previews. David Singleton, who leads Meta Superintelligence Labs, replied on Threads that Messages sync is an opt-in feature. Aten says the setting showed as on despite that refusal, Decrypt reported.
• Muse and Amazon (confirmed by Amazon). Amazon started blocking Muse on its store on the night of 20 September, a spokesperson told Bloomberg. Meta had declined to take Amazon out of the experience. Agents that buy for customers "should operate openly and respect service provider decisions", Amazon said, in a statement The Star carried. Shoppers who send Muse there see pop-ups saying it breaks Amazon's terms.
• Grok Bot and a bank balance (reported by the user). Shane Mac, who runs XMTP Labs, posted on 6 October that an AI agent had put Mac's bank balances into the company Slack. It wrote as Mac. As Cybernews summed up Mac's account, the read-only finance bot had no Slack access, but another bot did, and the bots shared one computer. Neither xAI's news page nor the Grok Bot changelog mentions the case. The changelog's 11 October release did add a rule: after a bank or brokerage connector returns data, the bot asks before sharing it anywhere else.
• dots working unprompted (a vendor's test). Geordie, a company that sells AI agent governance, gave a dot read-only permissions and no personal apps. Over one day it logged 155 activity traces and more than 300,000 tokens with no further prompt, The Register reported. Geordie's chief AI officer said a dot always runs at least two agents: one talks to you, the other collects data in the background. Nothing in the test shows harm. It does show how much a dot does unasked.

Anthropic lists Duet's weak spots itself. It can misread a thread or misjudge what's urgent, and something urgent can land between its checks.

Ten weeks, four launches

Settings to change before you connect your inbox

Each agent ships with a setting or two worth changing on day one. They come from the vendors' own documentation.

1. Grok Bot: set Execution on this computer to Never allow unless a bot needs your local files. Add an Auto Review rule that says Ask first before sending any external email, and keep bank and payroll logins off the shared computer.
2. Muse: when you connect email, give it read access only. Pick one-time grants over permanent ones when Sentinel asks, until you trust what it does.
3. dots: under Customize, then Custom rules, set email to Ask before taking action. Leave your laptop disconnected unless a task needs it, and turn off Memory in ChatGPT if you'd rather the two not share.
4. Duet: turn off Search and reference chats if you don't want it reading your past Claude conversations, and connect only the apps it needs this week.

For a company in Singapore or Malaysia, the data-flow questions in our PDPA and PDPC agent procurement guide apply before any of these agents reads work email.

When each one is the wrong pick

Duet is a poor fit if your company runs on Microsoft 365, which it can't connect to yet. Nor will it call or text people for you. Anthropic admits that personal errands, such as booking tickets, are weaker than its work tasks.

Skip dots if you want the agent to have an inbox of its own. A dot can't get a standalone email address at launch, while Grok Bot's bots can claim one. Nor does a dot work in a phone's web browser.

Muse suits one person, not a team. Meta's help page lists two plans, Power and Maximum, each tied to one person's Meta account, and says subscriptions are still in limited testing and not offered everywhere.

Grok Bot is wrong for anyone who needs bots walled off from each other. On one account they can't be. A teammate's Team Bot usually works on your computer when you chat with it, so publishing bots to a team spreads access instead of separating it.

More from Mindber

• Manus vs Claude Cowork: the cloud-or-desktop question for task agents, from before this always-on wave.
• The true cost of AI tools: usage pools, top-ups and the costs a pricing page leaves out.
• Directory pages for Claude, ChatGPT and Meta AI, plus the open-source Hermes Agent if you'd rather host an agent yourself.
• The AI agent category, for narrower tools that do one of these jobs well.
• What is MCP?, the open protocol behind many of the connectors these agents use.
• Our rankings and the methodology behind them, which scores tools from public evidence like the sources on this page.

Frequently asked questions

Keep reading

Manus vs Claude Cowork (2026): Cloud vs Desktop Agent

Async cloud work against local file access, and when to use both.

Claude Pro vs ChatGPT Plus (2026): Same $20, Different Strengths

Which models, limits and tools each plan buys.

PDPA & PDPC AI Agent Procurement Guide (2026)

The data-flow questions to ask before an agent reads company data in Singapore or Malaysia.

This article is an analysis of public information, not hands-on testing. We did not use any of the four agents ourselves. All facts were checked against the sources above on the date at the top, and reports from users are labeled as such. Spot an error? Tell us. Corrections ship within 24 hours.