Claude Duet Works While You Step Away, But People Still Approve Big Decisions
research preview
An early public version tested before a final release.
connector
A link that lets an AI use another app.
virtual machine
A computer created inside another computer system.
What happened
Anthropic, the company behind Claude, announced Claude Duet on October 21. It opened a research preview through a US waitlist for Pro and Max subscribers. The product works as a standalone app on iOS and Android. It also works through claude.ai.
Claude Duet is designed to act before a user asks another question. It sends a morning brief, handles assigned tasks, and contacts the user when something needs attention. It can keep working after the user leaves. A Hacker News post about the announcement received 329 points and 296 comments. Those numbers show community attention. They do not prove the product is correct or safe.
The background
Most chat assistants wait for people to open a conversation. Users then ask questions or give instructions. Claude Duet changes that timing. It carries over Claude memory, skills, and connected apps. People can message it by typing or voice, using one ongoing thread.
The system works inside a virtual machine with its own browser. The user’s laptop or phone does not need to stay on. Duet checks connected accounts every few hours, rather than constantly. Anthropic says constant checking would consume plan limits much faster. Background checks also count against the same limits as other Claude use.
Why it matters
The important change is not simply that the AI can answer questions. The change is that it can decide what deserves the user’s attention. It can prepare a reply, organize information, or build a document before the user asks for each small step.
Anthropic describes uses for professionals and small businesses. Duet can prepare meeting materials, find unpaid invoices, draft reminders, prepare payroll, and create dashboards. That could reduce the time spent moving information between many business tools. It also creates a new responsibility. An assistant that can act across many apps needs clear limits and visible decisions.
What is confirmed
Anthropic says Claude Duet can use more than 6,000 connectors and plugins. Its examples include Gmail, Slack, QuickBooks, Stripe, Gusto, DocuSign, HubSpot, and Shopify. The system can read information, prepare drafts, and create work for review.
Before sending a message, making a payment, signing a document, or deleting something, Duet shows the proposed action and waits for approval. Users can disconnect an app. Some actions are blocked even after approval. The announcement names securities trading and moving money between accounts.
The preview still shares usage limits with regular Claude. Larger tasks, such as financial models or slide decks, require more background work and model turns.
What remains unknown
Anthropic says the early version may misunderstand a thread or judge urgency incorrectly. It may miss something that arrives between checks. Personal errands are weaker than work tasks, and users cannot yet see the browser it operates.
The announcement does not provide independent success rates, error rates, or long-term evidence about how much correction users need. It also does not show whether frequent notices will feel helpful or distracting in daily work.
What to watch next
The next useful signals will be access beyond the US waitlist, possible Team and Enterprise support, and clearer evidence from real users. Watch how Duet explains its actions, records approvals, handles mistakes, and limits notifications. Claude Duet is not proof that an AI can run a business alone. It is a test of whether an AI can start useful work while keeping important choices with people.
Claude Duet Can Notice Work Before You Ask
📰 Full story: Claude Duet Works While You Step Away, But People Still Approve Big Decisions
Claude Duet watches connected apps and asks people to approve important actions.
research preview
An early version people can test.
approval
A person saying an action is allowed.
connector
A link between an AI and another app.
💡 The gist
- Anthropic, the company behind Claude, introduced Claude Duet.
- It checks connected apps and reports what needs attention.
- People approve important actions before Duet takes them.
What happened
Anthropic announced Claude Duet on October 21. It is a research preview. A research preview lets people test a product before its final release. Access begins with a US waitlist for Pro and Max subscribers.
Claude Duet can work as a separate app. People can also use it on claude.ai. It sends a morning brief about important tasks. It can check messages, calendars, invoices, and other business information.
A normal chatbot waits for a question. Claude Duet can start by noticing a problem. It might find an unpaid invoice. It might prepare a reply. It might organize information for a meeting. This matters because workers use many apps every day. Collecting the information can take longer than understanding it.
Duet works in a cloud computer. Your laptop does not need to stay on. It checks connected accounts every few hours. Checking constantly would use more plan limits. This explains why Duet is always available, but not watching every second.
Important limits
Anthropic says Duet can use more than 6,000 connectors and plugins. A connector links an AI with another app. Examples include Gmail, Slack, Stripe, and DocuSign.
Duet can read and prepare work. It asks for approval before sending messages. It asks again before paying, signing, or deleting something. People can disconnect an app. Duet cannot trade securities or move money between accounts, even with approval.
What is still unclear
The Hacker News post received 329 points and 296 comments. That shows attention, not accuracy.
Claude Duet is still early. It may misunderstand a conversation. It may choose the wrong urgent task. It may miss something between checks. Personal errands work less well than business tasks. Users also cannot see its browser yet.
Future testing should measure mistakes and corrections. It should also show whether notifications help people. The key question is simple. Can Duet save time without making people lose control? The answer is not known yet.
💬 Claude Duet: helpful assistant or another locked-in AI service?
Some commenters see Duet as a useful assistant for email, schedules, and routine computer tasks. Others worry about privacy, mistakes, and being tied to one company. The performance claims come from individual users, not independent testing.
- Duet is discussed as an always-on assistant like Dot or Muse. Some people think better Claude or Opus models would make it worth choosing; others want to switch between model providers.
- Users report delegating email, calendars, calls, shopping, and repetitive computer work to assistants.
- One user says a similar assistant helped with major car and home-buying processes. That is a personal account, not verified evidence.
- Another user says Dot repeatedly struggled with coding but was useful for finding important emails. Again, this is an anecdotal comparison.
- More access to email and other apps may not produce better answers; it may make an AI force connections that are not useful. Local AI can give users more control, but a commenter says even 24GB-VRAM local models are weaker than large cloud services.
- Some commenters say these assistants are less reliable than advertised. Others mainly debate whether Hacker News has too much AI news or whether that is simply the current technology trend.
initial digest at 296 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
Claude Duet Helps While You Are Away
📰 Full story: Claude Duet Works While You Step Away, But People Still Approve Big Decisions
Claude Duet is an AI helper that can keep working after you leave.
research preview
A version still being tested.
cloud computer
A computer used through the internet.
Hacker News
A website where people discuss news.
Anthropic, the company behind Claude, made Claude Duet.
Claude Duet is a helper like a small office assistant. It looks at connected email and calendars. It tells you about important things. It can keep working while you are away.
It uses a cloud computer. Your own computer can be turned off.
Claude Duet does not do every important action alone. It asks before sending messages. It asks before paying. It asks before signing papers. It asks before deleting things.
You can say yes or no. It cannot trade securities. It cannot move money between accounts.
Claude Duet is a research preview. That means it is still being tested. It can misunderstand a message. It can miss something new.
Hacker News is a website for discussing news. Its post got 329 points. It also got 296 comments.
That shows many people noticed it. It does not show that everything works correctly. People will learn more as testing continues.
💬 Claude Duet: a helpful robot with some worries
Some people think Duet could be a handy helper. Other people are worried. The performance stories come from people who used the tools themselves.
- Duet could help with email, plans, and repeated computer jobs. Some people want Claude because it may be smarter, while others do not want to be stuck with one company.
- Some users say similar helpers handled big shopping tasks or found important emails. Another user says coding went badly. These are personal stories, not proof that it will always work.
- Giving an AI lots of information does not always make it wiser. People also worry about private information; an AI on your own computer may be safer but weaker.
- Some people say AI helpers do not work as reliably as advertisements suggest. Hacker News users also argue about whether there is simply too much AI news.
initial digest at 296 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 Claude Duet: useful automation, but trust and lock-in remain open questions
The thread sees Claude Duet as a persistent assistant for email, scheduling, and repetitive computer work, while raising concerns about connected data, relevance, model lock-in, and reliability. Performance and failure claims are user reports, not independent tests.
initial digest at 296 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.