This Claude Duet review comes three days into using it, so treat it as a first, hands-on look rather than a final word. I've been peeling back the onion to see everything it can do, and the short version is this: it's the smartest of the work agents I've tested, and it's second in my ranking, behind OpenAI's dots. Given that it's still a research preview, I think that's the right spot for now.
Zen Aegis builds AI agent teams for businesses, so I test tools like this to know where they fit and where a custom system does better. Anthropic didn't pay for this review and there are no affiliate links in it.
What Claude Duet is
Anthropic's Claude Duet is a proactive agent, and it opened as a research preview on October 21, 2026. Access runs through a waitlist, starting with Claude Pro and Max subscribers in the US, and it comes as its own iOS and Android app plus claude.ai on the web.
You get one ongoing conversation that carries everything, with what really attempts to be a chief of staff. That's my phrase, not Anthropic's, but it's how it behaves: it brings your memory, skills and connected apps with it, sends you a short brief each morning, and checks your connected accounts every few hours. Under the hood it runs on Claude Opus 5.5, and you don't pick the model. I got in through the waitlist, on Claude Max.
Opus makes it smart, and you'll feel the cost
Running everything on Opus, Duet is pretty darn smart. When it does a job, it does it very well, and it takes on complex tasks better than anything else I've tested. I'll give credit where credit's due on that.
You pay for that intelligence in usage: Duet is heavy Opus usage, and it consumes quite a bit. Anthropic's own getting-started guide for Claude Duet says background work counts against your plan too, so you may reach your limits sooner. I did: even on Max, I hit my limit inside three days. That matters because of what comes next.
It tries too hard to be useful
My biggest frustration is how Duet starts. It comes with all the connections, and then it makes guesses at being useful. Briefs are turned on right away, and check-ins are turned on right away, without me asking and without it getting to know how it can be a good chief of staff for me.
I did take time to train it. I gave it context docs and websites and had it review all of it, and it was still guessing, trying too hard to be useful instead of understanding how to be useful. And I'm paying for those runs. I personally hold Claude and Anthropic to a higher bar, and I'd expect it to start by asking how it can help instead of guessing.
Channels are the best idea in Duet
If you take one thing from this Claude Duet review, make it channels, because this is where Duet does something nobody else does. It can create channels, and the chief of staff can manage them. If you run sales, marketing, ops and revenue, or you organize by customer, each one gets its own channel, and the chief of staff can go into each one and drive that work: CRM management, outreach, help with your website, whatever it might be.
The channels act like continuous sessions, and they share the same memory. Think about Claude Code, your other chats, maybe projects where you're trying to keep context down. I have so many Claude Code sessions in one repo, in one folder, in one project, and unless you challenge it with "Do you remember what happened in this chat?" it doesn't always make clear it knows. Duet's channels are meant to be persistent and long-running, and I love how the chief of staff keeps them going on its own, between my messages, so they take weight off my plate.
That's being thoughtful about a gap we've all hit in the ecosystem, and I give it kudos. As I play more with channels, I might fall more in love with them.
The approval problem: "I can't" versus "I need approval"
I recently asked Duet to clean up email in one of my connected inboxes, and it refused because deletes were involved. Instead of telling me it needed my approval, it said it can't do that.
That's tough to hear from a model that has access and is that smart. "I can't" and "I need approval" are very different answers, and the second one is what a good harness gives you. Anthropic's documentation says Duet is designed to ask before it deletes anything, so what I saw didn't match the design. Maybe I hit a category it won't touch at all; either way, it should have said so.
There's also no bypass-permissions option, and some steps aren't covered by a "don't ask me again" rule. Anthropic has given Claude Code an auto mode, so a Claude agent that won't even ask felt contradictory.
Any honest Claude Duet review has to weigh that against the safety argument. I'm sure an argument can be made about how I run my own systems on bypass permissions, but my system is good enough that I'm comfortable with it. I'm not asking Duet to give up security. I want policies: these actions are approved, these aren't, and this is when you ask.
Where Duet stops
These are the limits I hit or confirmed in three days. They're the little gaps you'd expect in a preview, but know them before you plan work around it:
- It has its own virtual machine with its own browser, but that isn't as obvious as it is in the other tools, and you can't watch it work.
- Unlike dots and Muse, it can't live on your computer and drive it. It can't see the files on your machine.
- It does quite a bit, but not infrastructure work like SSH or working with your VPS provider.
- It's built for work, and Anthropic says personal errands aren't as strong yet.
- Microsoft 365 isn't connected yet, and it doesn't make phone calls or send texts.
Claude Duet pricing, the waitlist and usage limits
Claude Duet pricing is simple during the preview: it's included in Claude Pro and Max at no extra cost. It isn't on Team or Enterprise plans yet. The catch is the Claude Duet usage limits, because Duet draws on the same plan limits as the rest of Claude, background work included.
Anthropic is softening that during the Duet research preview: new users get one full limit reset, usable until November 30, 2026. On the Claude Duet waitlist, access opens in waves, starting with US Pro and Max users. Max didn't keep me under the limit, so if you're on Pro and use Claude heavily for other work, plan for Duet to eat into it.
Claude Duet vs OpenAI dots and the others
Claude Duet vs OpenAI dots comes down to trust built over time versus raw intelligence. My dot, Victor, runs my whole business from one agent, lives on my computer when I need it to, and has earned my trust; my OpenAI dots review covers why it's first. Duet is smarter per task, but it guesses more and asks less well.
Each tool has its own identity. Duet isn't Grok Bot style, dots style or Muse's style. For your personal life, Muse is still the one I'd start with, and my Meta Muse review explains why. Grok Bot sits third; my Grok Bot review covers the hallucinations that put it there.
If you're choosing between Claude, ChatGPT and Gemini as everyday assistants before you get to agents, our guide to the best AI for business compares them. For ready-made AI employees instead of an agent, read my honest Marblism review.
Is Claude Duet worth it?
If you already pay for Claude Pro or Max, yes, get on the waitlist and try it: it costs nothing extra during the preview. The one rule from this Claude Duet review: set it up deliberately. Tell it how you want to be helped before it decides for you, and watch your usage in the first week.
I give Anthropic some grace here, because it's a research preview and I'm sure more is coming. It is an impressive one: it has gaps, but it also does some things better than anything else, and it's running on Opus, so you can't count it out. I hold Claude to a higher bar because their models sit near the top of the intelligence index, and when Duet comes out of preview, I expect it to go for the top spot, as they usually do. Three and a half stars out of five, for a preview.
Frequently asked questions
Is Claude Duet free?
It's included at no extra cost with Claude Pro and Max during the research preview. It isn't available on the free plan, and it uses your plan's limits, so it isn't free to run.
Can Claude Duet use my computer?
No. Duet works in its own virtual machine with its own browser. It can't see the files on your computer or drive it, which is a real difference from dots and Muse. If your work lives in local files or needs your own logged-in browser, plan around that.
What are channels in Claude Duet?
Channels are separate, ongoing threads for different topics, such as sales, operations or one customer, that share the same memory. The chief of staff can work across them and keep each one moving.
