Claude Duet vs OpenAI Dots vs Meta Muse vs Instinct vs Codot: Which AI Agent Should You Trust? (October 2026)

TL;DR:

• Claude Duet (Oct 21): bundled into Claude Pro at $20/month, so it's the cheapest way to get an agent on a top model. US-only and invite-only for now.
• OpenAI Dots (Sept 29): the most capable for long, ongoing goals. Costs $100+/month, and isn't available to Pro users in the UK or EU.
• Meta Muse (Sept 8): free, with a thoughtful payment design. It also read a journalist's private iMessages after he declined access.
• Instinct: the most autonomous, with the most aggressive data terms. Still invite-only.
• Poke: you just text it. The lowest-commitment way to find out whether you even want an agent.
• Codot (our product): you just talk to it. Daily briefings and scheduled automations like the big agents, built around voice, and free right now.

Six weeks ago, "personal AI agent" meant a handful of startups. Now Meta, OpenAI and Anthropic all have one. On October 21, Anthropic shipped Claude Duet, and the comparison we published two months ago already felt out of date.

The feature lists have mostly converged: background work, app connections, approval before anything risky. So I don't think "which one is smartest" is the useful question anymore. The useful questions are: which company do you trust inside your inbox, and how much effort does it take to hand it work?

Full disclosure: I build Codot, a voice-first assistant, so I've included it in the comparison. Weigh my take on it accordingly. I've tried to be as blunt about its gaps as about everyone else's.

Meta Muse: generous, and already in trouble once

On paper, Muse has the most careful design of the big three. Purchases go through Link by Stripe with single-use card numbers. Meta says the agent can't see your passwords and checks with you before it sends email or buys anything. And it's free for most people.

Then, two weeks after launch, a tech columnist noticed Muse referring to one of his private message threads, even though he says he had declined Messages access during setup. By the time he checked, it had synced a large part of his message history, and when he asked how it knew, the explanation Muse gave him turned out to be wrong. Meta says the feature was opt-in; he disputes that he ever turned it on.

This doesn't mean Muse is unusable. It does mean I'd connect it to accounts I could live with leaking, and check its settings screen after setup, not just during it. Early users are also split on the product itself, and the most common complaint in developer forums is that the underlying model feels a step behind the competition.

OpenAI Dots: the one people actually hand big jobs to

Dots are built for goals that take weeks, not minutes: "keep my inbox under control," "find me an apartment under $3,000." Each dot gets its own cloud computer that you can watch and take over, and you write rules for what it may do alone.

The most striking user reports so far are about scale: people handing a dot a whole multi-step purchase, like buying a car, and checking in only at the decision points.

The catch is the price. Dots only come with ChatGPT Pro, which starts at $100/month. There's no free or Plus-tier access, and Pro users in the UK, Switzerland and the EEA are locked out for now.

Claude Duet: the newest, and possibly the best deal

Anthropic announced Duet on October 21 as a research preview. It works like the others: it runs on its own virtual machine, keeps going when you close the app, gives you a short daily briefing from your connected apps, and checks before it acts. It also carries over the memory and connected apps from your existing Claude account, so setup should be light.

What stands out is the price. Duet has no separate charge: it's included in Claude Pro ($20/month) and Max. That makes it by far the cheapest way to get an agent running on a frontier model (Opus 5.5). The limits are real, though. It's US-only, you need an invite from the waitlist, and its background work counts against your plan's usage limits.

It's one day old, so there's no track record yet. In the launch-day discussion, the main argument for it was simply the model: if you already trust Claude more, the rest matters less.

Instinct and Poke: two very different bets

Instinct raised money at a $2.5B valuation in August by doing what the others won't, like reading a verification code out of your inbox to finish a booking. TechCrunch reported that its early terms claimed a "perpetual and irrevocable" license to content users connect. Some terms have since been revised. It's still invite-only. If you get in, give it a secondary email account, not your real one.

Poke is the opposite philosophy. It lives in iMessage, WhatsApp, SMS and Telegram, it has an actual personality, and Cognition bought it in July largely for that reason. It's not where you'd run a multi-week project. But if you're not sure you want an AI agent at all, texting one for a week is the cheapest way to find out.

What people actually use these agents for

Here's the part the launch videos skip. Under the big stories, everyday use is mundane. In most user discussions, the realistic job for Muse, Dots and Instinct comes down to email and appointments.

The hard part of "email and appointments" usually isn't doing them. It's that the thing never got written down in the first place. These agents need you to hand them a clear goal. A lot of life doesn't arrive as a clear goal.

It arrives like this:

• You come out of a two-hour workshop with three follow-ups promised, and you're already late for the next thing.
• Halfway through a run, you finally figure out how to fix the pricing page, and you have nowhere to put the thought.
• You're in the school pickup line when you remember the form, the dentist and the birthday gift, all at once.
• A supplier calls while you're driving and changes two delivery dates and a price.

These are the moments Codot is built for, and why I'd argue it belongs in this comparison rather than next to it.

Codot: the one you just talk to

Codot starts from your voice instead of a goal you've already typed out. Tap the lock-screen widget, or hold the button on your phone or Apple Watch, say what's in your head in whatever order it comes out, and Codot sorts it into calendar events, tasks, notes and contact details. It syncs with Google Calendar, so changes land where you already look.

It also does the "works while you're away" part that the big agents advertise, in a narrower and more predictable way:

• Daily briefings. Every morning Codot tells you what's on today, what's overdue, and the news you asked it to watch.
• Scheduled automations. Recurring jobs you set up once by voice, like "every Friday afternoon, summarize what I finished this week."
• A personal CRM. Mention that a client's daughter just started school, and it's there before your next call with them.

In App Store reviews, the thing people mention most is that it replaced a paper planner, and that events land in their phone calendar without being typed in one by one.

On trust, Codot only works with what you tell it and the calendar you connect. Your data is encrypted, and we don't sell it. On price, it's free during our launch period. After that, the free plan keeps 3 daily briefings and 2 scheduled automations, and Codot Max is $12.99/month ($119.99/year) for unlimited ones. That's a fraction of Dots at $100+, and you don't need to be in the US or on a waitlist.

Who Codot is best for

The big agents are built for people who can describe a goal and walk away. Codot is a better fit when your day doesn't hold still long enough for that:

• People with ADHD, or anyone who forgets things the moment they get distracted. Capture takes seconds from the lock screen, before the thought is gone, and the morning briefing brings it back when it matters. There's nothing to set up first.
• Founders and managers running several projects at once. One rambling voice note can touch three projects, two people and a deadline, and Codot files each piece where it belongs.
• People whose work happens away from a desk, like sales reps, real estate agents, consultants and store owners. Most of what you need to remember comes up in a car, a hallway or a client's living room, and you can log it from your watch.
• Parents running a household. The school form, the dentist and the grocery run all land in one calendar, not three apps.
• Anyone outside the US, or anyone who doesn't want to pay $20–$100 a month to find out whether an assistant is worth it.

Who should pick something else: if you mostly need errands done on the web, like booking, buying or form-filling, Dots or Duet will serve you better. If you run big projects with dependencies and Gantt charts, use a real project management tool. And if you're on Android, Codot is still waitlist-only.

If you have ADHD, our ranking of AI agents for ADHD by setup effort goes deeper on why low-setup tools tend to stick.

So which one?

• You already pay for Claude Pro and you're in the US: join the Duet waitlist. It costs nothing extra.
• You have a big, ongoing goal and $100+/month is fine: Dots, today.
• You want to try an agent for free: Muse, but connect it to a secondary account and recheck its permissions after setup.
• You're in the UK or EU: none of the big three is fully open to you yet. Poke and Codot work.
• Your real problem is that things never get written down, or you want this working today without a waitlist or a $100 bill: start with Codot. Add an execution agent later if you find yourself needing web errands done.

FAQ

Is Claude Duet free?

No separate fee, but you need a Claude Pro ($20/month) or Max plan, you need to be in the US, and you need an invite from the waitlist. Its background work counts against your plan's usage limits.

Claude Duet vs OpenAI Dots: what's the difference?

Both run on their own cloud machine and keep working while you're away. The practical differences are price and access: Duet comes with a $20/month Claude Pro plan but is US-only and invite-only, while Dots needs ChatGPT Pro at $100+/month and is generally available outside the UK, Switzerland and the EEA.

Is Meta Muse safe after the iMessage incident?

Meta says the Messages feature was opt-in; the journalist involved says he declined it. Until that's resolved, treat Muse like any new agent: connect secondary accounts first, and check which permissions are actually switched on after setup.

How is Codot different from Claude Duet or OpenAI Dots?

Duet and Dots are execution agents: you hand them a goal and they go do it on the web. Codot is built around speaking: it turns what you say into your calendar, tasks, notes and contacts, then keeps you on track with daily briefings and scheduled automations. It's free right now, available outside the US, and needs no waitlist. What it doesn't do is browse websites or make purchases for you.

If most of your to-dos start as a thought you're about to lose, download Codot free on the App Store or try it on the web. Say it out loud, and it becomes a task, an event or a note.