Claude Duet: The Most Cautious Agent in the Pack, With the Strongest Brain on the Market

This is Claude Duet, the personal agent Anthropic unveiled on October 21, 2026. It's the last of the big four to arrive: Grok Bot came out in August, Meta's Muse in early September, and OpenAI's dots at the end of September. It's the only one of the four that explicitly blocks financial actions, like trading securities or moving money between accounts, even if you approve them. And its model currently tops the independent Artificial Analysis index.

I read Anthropic's announcement, the help pages, the reviews from three people who got early access, and the Hacker News discussion, where one user who already has access also chimed in. I haven't used Duet myself: for now it's open only through a US waitlist. Throughout the article, I flag what's fact, what's reported and what's my opinion.

What is Claude Duet, and why the name?

Claude Duet is Claude that doesn't wait for you to type. In regular chat, you ask and it answers. In Duet, it starts things: it tells you each morning what needs your attention, keeps working while you're away, and comes to you when something's ready or when it needs a decision.

Anthropic didn't explain the name, but it fits: in a duet, each player carries half the piece, and neither plays the key note alone. Duet can read, research, draft and prepare on its own. When it's time to email a client, pay, sign or delete, it stops and shows you exactly what it's about to do. That principle explains almost every decision in the product, including the annoying ones.

Here's the video Anthropic launched it with:

Anthropic's Claude Duet announcement on X, October 21, 2026

A few things worth knowing up front:

• It's a separate app. Duet doesn't live inside the regular Claude app. It has its own app for iPhone and Android and, according to Anthropic, also runs on claude.ai. Anthropic says what it learns from the preview will shape how the two apps work together.
• It already knows you. Your memory, Skills and the apps you've connected to Claude carry over, so there's little to set up at the start.
• It's built for people who work. Anthropic pitches it to professionals and small business owners, and says itself that personal errands, like booking tickets, "aren't quite as strong."
• It's still an experiment. Anthropic calls it a "research preview." In plain English: it works, but expect it to change a lot.

What a regular day looks like

Big pieces of work come back to you as an editable file in the Make tab, and everything that's still open shows up in the Tasks tab. Only you can mark a task as done.

Tap through a day in the life of a small restaurant owner. This is an illustration I wrote based on what Anthropic describes, not a record of a real user.

Look at 11:40 AM. Duet didn't spot the double charge the moment it came in, but at the next check. Anthropic deliberately chose to check "every few hours" to save on usage limits, and says itself that it can miss something that arrives between checks. If you need an agent that responds to an urgent customer message within ten minutes, Duet isn't for you.

My take: more "no," and a stronger brain

Compared with its three rivals, Duet gives you less freedom and more judgment. Here's the price, and here's what you get for it:

Two things to know about this list. First: the one-time card for each purchase, and the fact that the agent never sees your passwords, exist in Muse too. They're strengths of Duet, but not advantages over Muse. Second: the 6,000 is Anthropic's number; the catalog itself currently shows about 1,000 connectors and about 340 plugins. A few smaller details sit alongside the limitations: Slack works with restrictions, there are no recurring payments through Link, you can't pick a model, and it doesn't call or text people. On the other hand, I know Opus 5.5 up close: it's the model I work with every day.

One point is worth getting exactly right, because it's easy to misread. Duet doesn't ask for more approvals than the others. The four actions that need approval (sending, paying, signing and deleting) are much the same as in Muse and dots. As VentureBeat noted, Muse also has approval gates and an isolated cloud computer. What sets Duet apart is the list of things it won't do at all, and the fact that it's less transparent about how it works under the hood.

Why the brain matters more in an agent than in a chat

In a chat, you see every answer and decide what to do with it. An agent works while you're not looking: it reads dozens of emails, decides what's urgent and drafts replies. If you have to double-check everything it does, you haven't saved anything. The real question is whether you can trust what it prepared without redoing the work.

A Hacker News commenter who has been using dots for a month put it exactly that way: a personal assistant running on a frontier model changes how much of its work you have to recheck. For him, dots handled a car purchase almost on its own, and is now helping with a home purchase in Japan. He wrote that he'd try switching to Duet because of Opus 5.5 as soon as he gets in. Another user in the same thread said Opus 5.5 found an $8,000 error in his tax return.

Here's how the models behind the agents score on an index that runs them all under the same conditions:

Two caveats, because they matter. First: Anthropic writes that Duet "decides how much effort a task needs," so not every answer gets Opus 5.5 at full strength. "Effort" is how long the model thinks before it answers. At high effort, Opus 5.5 scores 54, one point above dots, and at medium effort 51, below GPT-6 Astra at maximum effort. Muse currently runs on Muse Spark, and Meta has announced that Wonder is coming to it soon. The strongest version of Wonder, scoring 56, is available only in Meta's coding tool. Second: there is still no test comparing the four agents themselves head to head. A strong model is a precondition for a good agent, not proof that the agent is good. I wrote about how Opus 5.5 got to the top in Anthropic's comeback.

What early users are saying

Before launch, a handful of people got early access, including the business owners Anthropic quotes in its announcement. Three independent reviewers published reviews. They don't agree with each other, and that's the interesting part.

• Dan Shipper of Every. Two weeks of early access. He liked the per-topic channels, the running list of what the agent is doing and what it needs from him, and the suggestions "I actually want." He's missing a way to hook its scheduled runs up to his custom connectors and a true voice mode, and its email drafts are weaker than those from dots.
• Claire Vo, in Lenny's Newsletter. She tested email and Slack, a sales pipeline, a slide deck, a launch video and basketball tickets. She called it the most polished launch she's seen, but it never reminded her it existed, it's slow, and there's no desktop app.
• Peter Yang. Early access and a video review. He thinks the files it produces in the Make tab could be its edge.

How can the same product feel proactive to one person and forgettable to another? Both can be true. Duet checks in every few hours and sends an update only when there's something to say. If you expect a partner that talks to you all day, it'll feel like it vanished. If you want quiet, and to be contacted only when needed, it'll feel like someone is finally filtering for you.

Peter Yang on X: early access to Claude Duet, compared with Grok Bot, Muse and dots

And in a Hacker News thread of more than 300 comments, four things came up that aren't in the official announcement:

And the general reaction: some people don't get how it's different from the others. Yuchen Jin, former co-founder and CTO of Hyperbolic, joked that he already has ten personal agent apps on his phone, and there's room for one more. Anthropic is entering a crowded market.

Security and payments: how it buys without seeing your card

Anthropic devotes a big part of the announcement to security. The idea in one sentence: the agent holds the business's key ring, but the key to money and to actions you can't undo stays with you.

Here's how it breaks down:

The four pieces, in plain English:

• Passwords. Duet is built so that passwords and login tokens are stored outside the agent, so it never sees them. When a site asks for a password, you type it into a separate form on your phone, and it goes straight to Duet's browser without passing through the conversation. Sign-in through 1Password is coming soon.
• Action checker. Every Duet action also runs through the auto mode classifier from Claude Code, Anthropic's coding tool. It's a second layer, not a substitute for you.
• Sensitive-data filter. A new dedicated classifier, together with safety instructions, screens every outgoing message. It's built to block credit card numbers and ID numbers. Personal medical and financial information goes out only when you explicitly ask.
• Protection from malicious instructions. An email or a web page can contain instructions designed to trick the agent. That's prompt injection, and it's the biggest risk for any agent that reads email. Anthropic writes openly that no defense is perfect, and that this is one reason important actions require you.

What happens when Duet buys something

Duet pays through two digital wallets: Stripe's Link and Shopify's Shop Pay. Here's what a single purchase looks like:

One more detail worth knowing before something goes wrong: according to Anthropic, for orders and refunds you contact the store, for charges you contact Link or Shop, and when Duet picked the wrong product, store, quantity or option, you contact Anthropic itself. It's good to know in advance who to call when the agent gets it wrong.

What hasn't been published yet

Three things are missing, and you should know about them before connecting a business account:

On the rules, to be precise: you can tell Duet what to update you about and how often, but I found no way to exempt an action from approval or add an approval checkpoint of your own. In dots and Grok Bot, you can. And one more detail from the help page: when you disconnect an app, Duet stops accessing it, but remembers what it learned from it in earlier conversations.

Anthropic wrote that more safety details will be published in the coming days. When they come out, I'll update this article.

How much Claude Duet costs

Duet is included in a Claude subscription at no extra cost, at least during the preview: Pro at $20 a month, or Max. For comparison, dots requires a subscription of at least $100. That's the point Engadget highlighted: Anthropic arrived last, but getting in is cheap.

There's one catch. Duet draws on the same usage limits as chat and Claude Code: a limit for every five hours and a weekly limit. The background checks count too, and big jobs like a financial model or a slide deck use up more. If you already hit your limits working with Claude, you'll feel it. Anthropic hasn't published how much of your allowance a normal day with Duet uses. In dots it's the other way around: according to OpenAI, conversations with a dot don't count toward ChatGPT's usage limits. So the $20 vs. $100 comparison holds for the entry price, but not necessarily for how much work you'll get. Everyone let in gets one full reset of their limits, usable until November 30, 2026.

Claude Duet in Israel and in Hebrew

Not yet. As of October 22, 2026, Duet is open only through a waitlist, and only to Pro and Max subscribers with a US account. Invites go out in waves, with no guaranteed date, and Anthropic hasn't said when it will reach other countries or the Team and Enterprise plans. We write from Israel, but the same holds everywhere: outside the US, Duet isn't available yet.

There's no official word on Hebrew, and Anthropic hasn't published a list of supported languages for Duet. Claude itself works well in Hebrew, but that doesn't yet mean Duet will correctly understand a supplier's email in Hebrew or know how to draft a reply to an Israeli customer. We won't say it works until we've seen it. And here too, as with Muse, we don't recommend getting around the restriction: Duet requires a US Claude account, and paying through Link also requires a US payment method. Connecting your business accounts in a way that breaks the terms isn't worth it.

Claude Duet vs. Muse, dots and Grok Bot, in brief

Each of the four agents has a different home:

We updated the full comparison today, with the architecture, approvals, connections and a quiz to help you choose, in Muse, dots, Grok Bot or Claude Duet. On each one separately: Meta Muse, OpenAI dots and Grok Bot.

The market has changed: from an open-source project to four giants

A year ago, there was no such thing as a "personal agent" in today's sense. What has happened since explains both Duet and what will come after it.

Personal agentsNovember 2025 to October 2026

• Nov 2025 · A solo developer's open-source agent
• Jan 2026 · Clawdbot becomes OpenClaw
• Feb 2026 · Hermes Agent · OpenClaw's creator to OpenAI
• Apr 4 · Anthropic stops covering OpenClaw in its plans
• Aug 11 · Grok Bot
• Sep 8 · Muse
• Sep 29 · dots
• Oct 21 · Claude Duet

The category was born in open source, and with Claude's name on it. In November 2025, Austrian developer Peter Steinberger released an agent that runs on the user's own computer and takes tasks by message. In January it was called Clawdbot, a name Anthropic saw as infringing its trademark. He renamed it twice, to Moltbot and then to OpenClaw, and it became one of the most popular projects on GitHub: today it has more than 390,000 GitHub stars, the developer world's popularity signal. In February, Steinberger joined OpenAI and announced the project would move to an independent foundation. That same month, Nous Research released Hermes Agent, which now has about 256,000 stars.

Then the giants moved in. There were forerunners: Google unveiled Gemini Spark in May, and Microsoft built Scout directly on OpenClaw. But within ten weeks, from August 11 to October 21, four giant companies launched personal agents of their own. Meta didn't even hide where the idea came from: according to The Decoder, Meta's Nat Friedman said Muse was "heavily inspired" by OpenClaw, and Muse ships with a personality file, SOUL.md, nearly identical to OpenClaw's.

And Anthropic has its own story here. On April 4, it stopped letting Claude subscribers use their subscription through OpenClaw and similar tools, and started billing for that separately. Six months later, it's launching a personal agent of its own, inside the same subscription. From a business standpoint, it makes sense: the model provider wants the agent you live with every day to be its own.

What this means for you

Three things have changed, and I see all three when I set up agents for businesses:

The difference has moved to the brain that runs the agent, and to how far you can trust it. That's what makes lock-in matter: Duet runs only on Claude, dots only on GPT, and Muse only on Meta's models. In Grok Bot, Cursor picks the model, not you. Today Opus 5.5 leads, but it only came out on September 22, and in this race first place changes hands every few months. If you're tied to one company's agent, you're tied to its model too.

Meanwhile, open source hasn't gone away. Nous Research, which develops Hermes Agent, confirmed on October 7 that it raised $90 million at a $1.5 billion valuation and launched a business version, and in July OpenClaw moved to a nonprofit foundation backed by OpenAI, Microsoft, Nvidia, GitHub and Tencent.

With open source, the software is free, and you pay for the model and the server. Memory and files stay on your server, but whatever is sent to the model still passes through the model provider.

That last line, on security, is the crux. OpenClaw shows what happens when nobody takes on that responsibility: in February, security researchers found 341 malicious plugins in its plugin store, most of them installing software that steals passwords from Macs, and at the same time tens of thousands of installations were found exposed to the internet. Several vulnerabilities were found in Hermes too, one of them documented in GitHub's advisory database and fixed. Open source isn't safe or dangerous in itself. It's exactly as safe as whoever installs and maintains it.

Disclosure: Platin Digital is an OpenAI Select Partner, and we usually build our clients' agents on Hermes Agent, with whichever model fits each business best: from Anthropic, OpenAI, Google or others. When the right model is Opus 5.5, it goes into the business's agent without Duet. So Duet doesn't threaten me. It works exactly the way we do, an agent that prepares and waits for approval, and it isn't available in Israel yet.

Who Claude Duet is for, and who it isn't for

What to do in the meantime

Duet won't reach Israel tomorrow, but getting ready for it is the same as getting ready for any agent:

1. Write down three tasks that come back every week and keep getting pushed off. Chasing invoices, replying to suppliers, a weekly summary. That's the first task list for any agent, Anthropic's or anyone else's.
2. Decide what always comes back to you for approval. Payments, messages to customers, publishing and budgets. Duet decided this for you. In your own agent, you decide.
3. Sort out permissions. Who holds access to email, bookkeeping and the website. A good agent starts with the minimum access it needs.

If you don't want to wait, you can set up an AI agent for your business that works from Israel today, with Claude or another model, in a channel you already use. We laid out what goes into a setup like that in our guide to building an AI agent. And when Duet opens here, we'll test it ourselves and write about whether it's enough for you without anyone in the middle.

Frequently asked questions

Sources

• https://claude.com/resources/articles/claude-duet
• https://claude.com/product/claude-duet
• https://support.claude.com/en/articles/17631495-get-started-with-claude-duet
• https://support.claude.com/en/articles/17666484-how-to-use-link-by-stripe-with-claude-duet
• https://support.claude.com/en/articles/17666489
• https://support.claude.com/en/articles/12138966-release-notes
• https://www.anthropic.com/engineering/managed-agents
• https://www.anthropic.com/engineering/claude-code-auto-mode
• https://www.anthropic.com/research/trustworthy-agents
• https://www.anthropic.com/research/prompt-injection-defenses
• https://claude.com/marketplace/connectors-plugins
• https://x.com/claudeai/status/2112984664256241837
• https://x.com/claudeai/status/2112984669943968099
• https://x.com/_catwu/status/2112988751576330487
• https://x.com/felixrieseberg/status/2112993956611178996
• https://x.com/petergyang/status/2112995494612082931
• https://x.com/Yuchenj_UW/status/2112985163764466145
• https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent
• https://www.lennysnewsletter.com/p/claude-duet-is-beautiful-i-still
• https://news.ycombinator.com/item?id=50186037
• https://venturebeat.com/technology/anthropics-answer-to-muse-is-here-claude-duet-a-mobile-and-desktop-ai-agent-that-takes-actions-proactively
• https://www.engadget.com/2293646/claude-duet-is-anthropics-answer-to-meta-muse-and-openai-dots/
• https://9to5mac.com/2026/10/21/anthropic-takes-on-chatgpt-dots-and-meta-muse-with-claude-duet/
• https://www.macrumors.com/2026/10/21/anthropic-claude-duet/
• https://artificialanalysis.ai/leaderboards/models
• https://techcrunch.com/2026/10/21/meta-says-muse-has-over-10m-users-with-shopping-a-top-activity/
• https://the-decoder.com/metas-ai-agent-muse-draws-500000-users-in-a-week-along-with-claims-it-copied-openclaw/
• https://en.wikipedia.org/wiki/OpenClaw
• https://steipete.me/posts/2026/openclaw
• https://www.infoworld.com/article/4154435/anthropic-cuts-openclaw-access-from-claude-subscriptions-offers-credits-to-ease-transition.html
• https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html
• https://github.com/openclaw/openclaw
• https://github.com/NousResearch/hermes-agent
• https://hermes-agent.nousresearch.com/docs/
• https://techcrunch.com/2026/10/07/nous-research-confirms-it-hit-1-5b-valuation-launches-ai-agents-for-business-users/
• https://github.com/advisories/GHSA-pgp4-xr4j-h5cg
• https://conscia.com/blog/the-openclaw-security-crisis/
• https://www.forkable.io/p/openclaw-is-now-a-non-profit-foundation
• https://thelettertwo.com/2026/06/07/openclaw-microsoft-google-meta-ai-agents