# Claude Duet Claude Duet is [[Anthropic]]'s proactive agent, announced on 21 October 2026 as a research preview. Built on [[Claude Opus 5.5]] and running on its own cloud computer, it watches your connected apps, flags what needs your attention, takes on the tasks you hand it, and keeps working while you're away. Before acting on your behalf, it asks. Anthropic's tagline: "Get back to what needs you." The pitch is the difference with the regular [[Claude]] app. The Claude app waits for you to open a chat. Duet "comes to you": every morning it sends a short brief on the most important items across your connected apps, and it messages you when work is done or when it needs a decision. Anthropic aims it at "professionals and small business owners with more on their plates than they can track". Thurrott notes that the focus is on work rather than personal errands. It's Anthropic's answer to [[Meta Muse]] and OpenAI's [[Dots (AI)|Dots]] (launched at DevDay on 29 September 2026), and it arrives in a category that self-hosted assistants like [[OpenClaw]] and [[Hermes Agent]] opened first. ## How it works - **One thread, text or voice.** You message Duet "like a teammate you trust". No prompts, no setup. The same ongoing thread handles quick requests and weeklong projects - **Its own cloud computer.** Duet works in a virtual machine with its own browser, so your laptop and phone can be off - **Check-ins every few hours.** It checks your connected accounts every few hours. According to Engadget, that interval is there to conserve usage limits - **Finished work comes back for review**, as docs, sheets or decks. You can see what it has done and what it's doing, and steer it along the way - **You set the noise level.** Per Engadget, you choose how often it messages you and about which topics - **Your Claude account carries over.** Memory, skills and connected apps work from the first message ## What it can do Anthropic's examples: - Scheduling meetings - Building dashboards that draw on dozens of tools (the product page shows "Marlo Pulse", a daily business dashboard) - Financial or sales forecasts, prototypes, web apps - Finding unpaid invoices and drafting the reminders - Preparing payroll for your review - Getting contracts ready to send - Morning briefs on hiring, contracts, permits, investor prep and customer issues The testers quoted at launch include Anthropic staff ([[Boris Cherny]], Chelsea Larsson and Karan Ahuja, for to-dos, scheduling and product work) and a few outside users. Jess Provencher, in real estate, says it caught a misscheduled showing. Amin Ebrahimi uses it for accounting, support and marketing across several restaurant locations. Mya Dalton, a brand strategist, got a call flagged before her flight: "I love that proactivity." ## Connectors and purchases Duet gets access to more than 6,000 connectors and plugins from the [[Claude Marketplace]]. The ones Anthropic names: Gmail, Slack, Google Drive, Notion, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva and Shopify. It can also make business purchases, with your approval, through Stripe Link (a virtual card) and Shopify Shop Pay. Sign-in through 1Password is announced for later. ## Safety model - **Approval before consequential actions.** Duet shows the action it proposes and waits before sending messages, making payments, signing documents or deleting anything - **Hard limits.** Trading securities and moving money between accounts are off-limits, even if you approve - **No credentials in the model.** Duet never sees passwords or sign-in tokens. The credential design follows [[Claude Managed Agents]], and approvals, logins and safety checks run inside your own cloud computer - **A classifier reviews actions.** Approvals go through the auto mode classifier used in Claude and [[Claude Code]] (see [[Claude Code Auto Mode]]), plus [[Prompt injection]] defenses. Anthropic cites its trustworthy-agents framework - **Strictest data tier blocked.** A separate classifier and the system prompt block sharing credit card numbers and government ID numbers. Health and personal-finance information is only shared on your explicit instruction - **Tested from outside.** Red-teamed, external penetration test, covered by Anthropic's public bug bounty. More safety details were promised "in the coming days" - **Disconnect anytime.** You can remove any connected app Anthropic's line on it: "We designed Claude Duet to be a partner you can trust." See also [[Human-in-the-Loop]]. ## Availability and pricing - **Waitlist, US only**, for Claude Pro and Max subscribers. Team and Enterprise are "not included yet" - **No separate price.** Duet shares your plan's usage limits with the rest of Claude, [[Claude Code]] included. Background checks count, and bigger tasks (a financial model, a deck) take more model turns - **Its own app** on iOS and Android, plus claude.ai on the web. During the preview, it lives apart from the main Claude app - Engadget says Pro ($20 per month) subscribers get access when Duet reaches general availability. Anthropic's page lists both Pro and Max on the waitlist, so sources disagree on whether Pro gets in during the preview ## Limitations Anthropic lists these itself: - "It may misread a thread or miscalibrate what's urgent." - "It can miss something that arrives between checking in with you." That's the price of checking every few hours - Personal errands, like booking tickets, are weaker - You can't watch its browser yet And a few things Anthropic hasn't said: whether you can view or edit memory from inside Duet (it reuses your Claude memory, which you can view and edit in claude.ai settings), how its data is used for training, and when regions beyond the US follow. ## How it compares Six big companies now sell a hosted, always-on agent. Two self-hosted projects started the category. Eight columns in one table is unreadable, so here are two. ### Hosted agents | | Claude Duet | [[Dots (AI)\|Dots]] | [[Meta Muse]] | [[Grok Bot]] | [[Gemini Spark\|Gemini Task]] (ex-Spark) | [[Microsoft Autopilot]] | |---|---|---|---|---|---|---| | Maker, launch | [[Anthropic]], 21 Oct 2026 (research preview) | [[OpenAI]], 29 Sep 2026 | Meta, 8 Sep 2026 | [[xAI]] (now SpaceXSI), run by [[Cursor]]; beta 11 Aug, enterprise 3 Sep 2026 | [[Google]], 19 May 2026; renamed Task in Oct | [[Microsoft]]; Scout 2 Jun, Autopilot 25 Sep 2026 (private preview) | | Hosting | Anthropic cloud computer (VM and browser) | OpenAI cloud computer per dot; your computer if you allow it | Dedicated VM per user (Muse Secure VM); Mac app for local files | One shared cloud computer per member, in the US; one screen per Bot | Google Cloud VMs; your local Chrome on desktop; linked Mac folders | Microsoft cloud, inside your tenant (Scout was a desktop app) | | Model | [[Claude Opus 5.5]] | [[GPT-6 Astra]] | Muse Spark (Muse Wonder announced for subscribers) | Router: xAI and third-party models (Opus 5.5 named), no picker | Gemini 3.5 at launch; model selectable per task now | Not stated (Scout used GitHub Copilot's catalog) | | Entry price | Claude Pro ($20) or Max; no extra charge, shares your plan's limits | ChatGPT Pro ($100+) or Business Premium | Free; Power $20; Maximum $100 | X Subscription Premium $30 (up to $200), or a paid Cursor plan | Google AI Pro $19.99; Ultra $99.99+ | Usage-based billing, no prices published | | Channels | Duet app (iOS, Android), claude.ai | ChatGPT, Slack, Teams, texting beta | Muse apps (iOS, Android, Mac, iPad), muse.ai, WhatsApp | Desktop (macOS, Windows, Linux), iOS, Android, Slack, its own email address | Gemini app (mobile, Mac, web) | @mention in Teams, Outlook, chats and documents | | Proactivity | Morning brief; checks your apps every few hours | Read-only background research; scheduled work | Unprompted suggestions; keeps working toward your goals | Routines on schedules or events; proactive primary Bot since 2 Oct | Schedules and event triggers you define; Daily Brief is a separate product | Watches channels, follows up, runs recurring work (Scout's heartbeat: every 15 to 120 min) | | Approvals and credentials | Approval before messages, payments, signatures, deletions; no securities trading or transfers even with approval; never sees passwords | Action rules plus a separate Auto-review model; passwords never enter the model's context | Sentinel decides every outbound action (allow, deny, ask); the agent only sees stand-in tokens | Bots act as you; Auto Review; approval before irreversible actions; tokens kept server-side | Review before sending, buying, forms, shared edits; passwords and payments handed back (Chrome's saved passwords with permission) | Each agent has its own Entra identity; org policies, Purview, audit logs | | Memory | Your Claude memory carries over; Duet-specific controls not documented | ChatGPT memories plus its own; you can't view or edit individual ones | Markdown files (`MEMORY.md`) you can open; Forget skill (TIME: may not fully forget) | Per-Bot memory; remember or forget by asking; no viewer documented | Personal Intelligence and past chats; no separate agent memory | Memory across conversations (Scout stored history in OneDrive) | | Region | US only (waitlist) | Pro excludes the EEA, Switzerland and the UK; Business Premium in all ChatGPT regions | US and Canada | Computers in the US; X Subscription US first; EU unverified | Everywhere Gemini is, except the EEA, Switzerland, the UK and Nigeria | No country list published (Scout: Frontier enrollment) | ### Self-hosted agents | | [[OpenClaw]] | [[Hermes Agent]] | |---|---|---| | Maker, launch | [[Peter Steinberger]], now the OpenClaw Foundation; first published November 2025, 2.0 in August 2026 | Nous Research | | Hosting | Your Mac, Windows or Linux machine, or a server you control | Local, Docker, SSH or serverless backends; a $5 VPS works | | Model | 35+ providers | Any provider (OpenRouter, Nous Portal, OpenAI, Grok OAuth...) | | Entry price | Free (MIT), plus your model bills | Free (MIT), plus your model bills | | Channels | WhatsApp, Telegram, Slack, Discord, Signal, iMessage and many more | Telegram, Discord, Slack, WhatsApp, Signal, Email and 15+ platforms | | Proactivity | Automations, heartbeats and hooks you configure | Cron scheduler; a learning loop that writes its own skills | | Approvals and credentials | Sandboxing and execution approvals off by default; approvals bound to the exact request since 2.0; the Secret Store isn't encrypted at rest | You set it up; credential firewall for Docker sandboxes | | Memory | Plain Markdown files in the workspace, with a SQLite index | Agent-curated facts, full-text search over past sessions, user model | | Region | Anywhere | Anywhere | ### Duet came last Google went first with Spark (19 May), then Grok Bot (11 August), Meta Muse (8 September), Microsoft Autopilot (25 September) and Dots (29 September). Duet arrived on 21 October, five months after Spark. Thurrott's verdict: Anthropic was slow to enter. They don't all aim at the same people: - **Duet**: professionals and small business owners - **Dots**: ChatGPT power users and teams already in Slack or Teams - **Muse**: consumers first (the only free tier), with a small-business version since 29 September - **Grok Bot**: teams, sold through Cursor to engineering orgs, then spread to sales, recruiting and finance - **Gemini Task**: people whose life runs on Gmail, Calendar and Drive - **Autopilot**: companies on Microsoft 365, enterprise-only ### Where Duet is better, where it's worse - **vs Dots**: same idea, a fifth of the entry price ($20 against $100+), and memory you can at least view and edit on claude.ai. Dots goes where teams already work (Slack, Teams) and lets you open its computer to watch or take over; you can't watch Duet's browser yet - **vs Muse**: Duet has harder limits on money (no transfers, no securities, even with approval) and none of Meta's privacy baggage (FTC history, Cambridge Analytica, TIME's dossier report). Muse is free to start, goes where people already are (WhatsApp), publishes a much more detailed isolation design (Sentinel, stand-in tokens) and lets you open its memory as Markdown files. Both pay through Stripe Link and Shop Pay, with 1Password announced - **vs Grok Bot**: Duet keeps each user in a cleaner box. Grok Bots share one computer and act as you, which is how one user's balances ended up in his company Slack. Grok Bot has far more channels (Linux desktop, Slack, email), rosters of Bots that work together, and model routing; Duet is one agent, one model - **vs Gemini Task**: Duet is more proactive out of the box (a morning brief and regular check-ins, where Task mostly runs the triggers you define). Task starts with deeper first-party data (Gmail, Photos, Maps, Chrome with saved passwords) and costs about the same ($19.99 for AI Pro) - **vs Autopilot**: different buyer. Autopilot gives each agent its own directory identity and lives inside Teams and Outlook, which is better for governance. Duet works for a solo professional with a $20 plan; Autopilot needs an employer with Microsoft 365, a Copilot setup and a preview invite - **vs OpenClaw and Hermes**: see below ### From Belgium None of the consumer offers works here today. Duet is US only, Muse is US and Canada, and Gemini Task and Dots on Pro exclude the EEA explicitly. Grok Bot runs its computers in the US and rolls out The X Subscription US first; nothing excludes Europe, nothing confirms it either. Autopilot publishes no country list, but it's an enterprise preview. The one documented route is Dots through ChatGPT Business Premium, which OpenAI lists in all supported ChatGPT regions. For an individual in Belgium, self-hosted (OpenClaw or Hermes) is the only option that works today. ### Managed or self-hosted Duet and Dots are the same bet: a managed agent with a cloud computer, tied to a subscription. The interesting comparison is with [[OpenClaw]] and [[Hermes Agent]], because that's the real choice: let a vendor run the agent, or run it yourself. What you get with Duet: zero setup, a credential design where passwords never reach the model, an approval classifier and hard limits you didn't have to configure, and a security team that red-teams the whole thing. OpenClaw's defaults are the opposite (sandboxing and execution approvals are off, and the baseline assumes a single trusted operator), so you have to build that safety yourself. Microsoft did exactly that for Autopilot, on top of OpenClaw. What you give up: model choice (Claude only), channels (Duet's app and the web, no WhatsApp or Telegram), and control over where your memory lives. OpenClaw keeps memory in plain Markdown files on your machine. Hermes stores its facts and session history on your hardware too, and both import each other's memory. With Duet, memory is your Claude account's memory. And the cost model differs: Duet costs nothing extra but draws from the same usage limits as everything else in Claude, while the self-hosted options are free software with API bills that grow with use. Proactivity is closer than it looks. Duet's morning brief and every-few-hours checks are a managed version of what OpenClaw users set up with heartbeats and automations. The self-hosted version can check more often, if you're willing to pay for the tokens. ## My take I live in Belgium, so for now this is a product I read about. US only, on a waitlist. The safety design looks right. Passwords that never reach the model, a classifier on every consequential action, and a short list of things it won't do even with your approval. Those are exactly the guardrails OpenClaw users have to add by hand. Memory is where I hesitate, as with Dots. Duet reuses your Claude memory, which at least you can view and edit on claude.ai; that's better than a dot, where you can only reset the whole thing. But I keep my agents' memory in plain Markdown inside my vault, so I can read it, fix it and take it anywhere ([[AI Agent Memory]], [[AI Agent Portability]]). And I already run my own agent system on top of [[Claude Code]]: agents with their own identity, memory files and daily logs, all in my notes. Duet would run beside that system, with its own memory in my Claude account. The usage limits bother me more. Duet shares your plan's budget with Claude Code, and its background checks count. If Claude Code is where you do your real work, an agent that polls your inbox every few hours spends from the same pot. And then there's the product story. In September, Anthropic merged [[Claude Cowork]] into the main Claude app because, according to [[Cat Wu]], people didn't want to choose which Claude product to use. Claude now keeps working after you close the laptop. About a month later, Anthropic ships a proactive agent... as a separate app. The New Stack had even titled its 3 October piece "Anthropic's answer to Dots and Muse is already inside Claude". Maybe the split is temporary (Anthropic says Duet lives apart "during the preview"). For now, it's one more Claude to pick from. ## Caveats - Features, examples, tester quotes and safety claims come from Anthropic. None of it has been tested independently yet - Duet launched two days before this note was written. It's a research preview: pricing, limits and regions will change - Engadget's claims (the polling interval being about usage limits, message frequency settings, Pro access at general availability) aren't on Anthropic's page - Anthropic's Duet page says 6,000+ connectors and plugins; my [[Claude]] note had the Marketplace at more than 2,000. Either it grew a lot or the two counts measure different things - The comparison tables come from my notes on each product ([[Meta Muse]], [[Grok Bot]], [[Gemini Spark]], [[Microsoft Autopilot]], [[Dots (AI)]], [[OpenClaw]], [[Hermes Agent]]), written from vendor pages and press between May and October 2026. All of these products change monthly - Grok Bot: xAI's marketing says each Bot has its own computer; Cursor's docs and the x.ai FAQ say all of a member's Bots share one. The table follows the docs - Microsoft Autopilot's requirements, model, prices and data-residency terms aren't published yet; several cells describe Scout, its predecessor - Meta's help center publishes weekly token quotas for Power and Maximum; the free quota isn't published. Muse Wonder is announced for the Muse app, but no report shows the switch yet - Gemini Task: Google calls it proactive through schedules and triggers; that it doesn't scan your apps unprompted is my reading of its docs - Region cells reflect what each company publishes. Where nothing is published (Grok Bot in the EU, Autopilot), that's unknown, not available - I found no Hacker News thread about Duet as of 23 October 2026 ## References - Product page: https://claude.com/product/claude-duet - Announcement: https://claude.com/resources/articles/claude-duet - CNBC: https://www.cnbc.com/2026/10/21/anthropic-preview-of-claude-duet-personal-ai-agent-take-on-muse-dots.html - Engadget: https://www.engadget.com/2293646/claude-duet-is-anthropics-answer-to-meta-muse-and-openai-dots/ - 9to5Google: https://9to5google.com/2026/10/21/claude-duet-anthropic-ai-agent-waitlist/ - MacRumors: https://www.macrumors.com/2026/10/21/anthropic-claude-duet/ - Thurrott: https://www.thurrott.com/a-i/anthropic/343121/anthropic-announces-claude-duet-a-personal-ai-agent - TestingCatalog: https://www.testingcatalog.com/claude-duet - The New Stack, "Anthropic's answer to Dots and Muse is already inside Claude" (3 October 2026): https://thenewstack.io/claude-answer-to-dots-muse/ - Meta Newsroom, "Introducing Muse" (8 September 2026): https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ - Meta AI Research, "How We Built Safety Into Muse": https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse - Meta Help Center, Muse subscriptions: https://www.meta.com/help/subscriptions/1021145227643680/ - Grok Bot page and FAQ: https://x.ai/bot - Cursor docs, Grok Bot security: https://cursor.com/docs/grok-bot/security - Techdirt, Grok Bot bank-data leak (19 October 2026): https://www.techdirt.com/2026/10/19/whoops-grok-bot-spills-tech-ceos-bank-data-in-public-slack/ - Gemini Help, "Task (formerly Spark)": https://support.google.com/gemini/answer/17094507 - Microsoft, "Introducing the new Copilot with Home, Code and Autopilot" (25 September 2026): https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/ - OpenClaw blog, "Microsoft Autopilot is built on OpenClaw" (25 September 2026): https://openclaw.ai/blog/microsoft-autopilot-openclaw - OpenAI, "Getting started with your dot": https://help.openai.com/en/articles/20001530 ## Related - [[Anthropic]] - [[Claude]] - [[Claude Opus 5.5]] - [[Claude Cowork]] - [[Claude Code]] - [[Claude Code Auto Mode]] - [[Claude Managed Agents]] - [[Claude Marketplace]] - [[Claude Tag]] - [[Dots (AI)]] - [[OpenClaw]] - [[Hermes Agent]] - [[Meta Muse]] - [[Grok Bot]] - [[Gemini Spark]] - [[Microsoft Autopilot]] - [[OpenAI]] - [[Google]] - [[Microsoft]] - [[xAI]] - [[Cursor]] - [[Peter Steinberger]] - [[AI Agents]] - [[AI Assistants]] - [[AI Agent Harness]] - [[AI Agent Memory]] - [[AI Agent Portability]] - [[Human-in-the-Loop]] - [[Prompt injection]]