Claude Duet: What It Is and How It Differs from Claude Code Routines

The sources here are Anthropic's product page and announcement, plus one Hacker News thread. Anthropic has not published pricing beyond "included in Pro and Max", and says more detail on safety design is coming. Nothing below comes from hands-on use of Duet.

If you searched "claude duet": it is a new Anthropic app, opened on 21 October 2026 as a research preview, where Claude acts as a proactive agent. It briefs you each morning, keeps watching your connected apps, and messages you only when something needs a decision. Access is a waitlist for Pro and Max subscribers in the US.

What Anthropic says it is#

Anthropic's announcement and the product page describe a few concrete pieces:

• Model: built around Claude Opus 5.5.
• Where it runs: a virtual machine with its own browser, so it keeps working when your laptop or phone is off. You get a notification when a task finishes or needs review.
• Cadence: it checks your connected accounts "every few hours" rather than nonstop, because continuous checking would burn through plan limits faster. You can tell it what to report and how often.
• Context: it reuses the memory, skills and connected apps from your Claude account, in one ongoing thread.
• Reach: Anthropic says it works with over 6,000 connectors and plugins from the Claude Marketplace, and names Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, Docusign, Canva and Shopify among the examples.
• Approvals: before it sends a message, pays, signs a document or deletes anything, it shows you the action and waits. Trading securities and moving money between accounts are off-limits even with approval.
• Safety stack: actions pass through the auto mode classifier used in Claude Code, credentials are handled the way Claude Managed Agents handles them so Duet never sees passwords or tokens directly, and a new classifier is meant to block sharing the strictest tier of data such as card numbers and government IDs.

Anthropic is also candid about limits: it may misread a thread or miscalibrate what is urgent, personal errands are weaker than work tasks, it cannot yet show you its browser, and it lives in its own app separate from the main Claude app.

The angle: Anthropic is moving the agent from your terminal to your inbox#

The Stratechery move is to ask what the packaging changes. Claude Code made the agent something a developer starts. Duet makes the agent something that starts the conversation. That flips the product from "pull" to "push", and it changes the economics: the value is in judgment about when to interrupt, and the cost is background usage that never shows up as a visible chat.

Who wins. Owners of small businesses and busy professionals who will never write a cron expression. The examples in the announcement are a broker, a restaurant owner and a brand strategist, not engineers. Anyone already paying for Pro or Max also gets it without a new line item during the preview.

Who loses. Single-purpose assistant apps that sell triage and scheduling on top of a model, since the model vendor now ships the same thing with your existing memory and connectors attached. Developers should note the connector count is the real moat: 6,000 integrations is hard to match with a weekend project.

Second-order effect. Plan limits become a product surface. Anthropic says background checks count against the same limits as chat. If proactive agents become common, expect usage-limit anxiety to move from "I ran out mid-session" to "my assistant used it overnight". The cost side of running agents unattended is covered in the overnight bill and agent FinOps.

What a Claude Code user does Monday#

There is no verified setup command to give you, because Duet has no developer API or CLI in the sources I read. What is worth doing:

1. Join the waitlist only if you are on Pro or Max in the US. Nobody else can use it yet.
2. Compare it to what you can already build. Routines and managed agent schedules cover repeating work against your repos, and Cowork covers desktop file work. Duet is the non-developer version of the same idea.
3. Treat a mailbox-reading agent as a security decision. Anthropic describes pen testing and a bug bounty, but the safety write-up it promised has not shipped yet, so wait for it before connecting a work account.

What people are actually saying#

The Hacker News thread had several hundred comments at the time of writing. Paraphrased:

• "This is just routines with a bow on it." Several commenters argued that Claude Code routines plus a long-running session can already triage email on a schedule, and that Duet packages existing tools for people who have not set them up. One said they had run sessions for days with a state file and a watcher script.
• The counter-case. Another commenter replied that a proactive agent reacts to events and keeps long-running task memory, while an ordinary session waits for input and compacts. That is the distinction Anthropic itself draws between Duet and the Claude app.
• Speed. One commenter relayed a video preview that found Duet slow to use even if smart, and argued that speed matters for an assistant that has to feel connected. This is secondhand and I did not watch the video.
• Trust. One commenter said they would not give Anthropic this much access to their life yet and would trust a company with a longer track record on mail and calendar more.
• Economics. A commenter suspected proactive agents are high-margin for labs because background work can fill idle inference capacity at low priority. That is speculation, not something Anthropic has said.
• Competitors. Commenters compared it with other proactive assistants, with mixed reports: one found one of them pesky about emails that did not matter, another liked its VM. These are single-user anecdotes.

Continue Reading#

• Claude Code routines vs managed agent schedules - the developer version of scheduled agent work
• Claude Cowork vs Claude Code - where each agent fits
• Claude Code auto mode explained - the classifier Duet reuses for approvals
• Claude Opus 5.5 release guide - the model under Duet

Sources#

• Anthropic: Introducing Claude Duet - 21 October 2026
• Claude Duet product page
• Hacker News discussion