Every other AI tool waits for you to ask. Claude Duet sends you a brief at 7 AM: a store can't print receipts, the vendor quietly pushed the fix to Monday, and two of your meetings clash. The escalation is already drafted, and nothing goes out until you tap Approve. Here's what people actually use Duet for, and a 15-minute demo you can run on made-up data.
Irv CassioReference guideEveryday users · IT & store support · presentersAbout 14 min read
Claude Duet is Anthropic's proactive agent app, in research preview since October 21, 2026. It runs on Opus 5.5 in its own cloud computer, sends a short morning brief, checks your connected apps every few hours, and asks before it sends, pays, signs or deletes. Reviewers keep coming back to the same five jobs: the brief, the unprompted catch, "who's waiting on me?", chasing money, and meeting prep. The best demo is one morning brief with three planted catches and one approval tap, all on synthetic data, because Duet has no Microsoft 365 support and no Team or Enterprise plan yet. Keep it personal, keep it synthetic, and read every approval card.
Duet isn't a chat window with a new name. It's a separate app with its own computer in Anthropic's cloud, and it starts the conversation.
Figure 1 · The loop Anthropic describes for the research preview (October 21–23, 2026)
Claude Duet facts as of October 30, 2026
Fact
Detail
Launched
October 21, 2026, as a research preview. Waitlist only, for US Pro and Max subscribers. No separate price.
Who's getting in
On October 28, Duet product lead Karan Ahuja said Anthropic is "getting all Max users in the US off the waitlist first." US Pro comes next. No date for international access, and no waitlist numbers.
Model
Claude Opus 5.5 only; Duet picks the effort level. It runs on its own cloud computer with its own browser, so it keeps working with your phone off.
The loop
A short morning brief, then checks of your connected apps every few hours. It messages you when something needs you.
Channels
One ongoing thread, plus topic channels that share the same memory. Your Claude memory, skills and connected apps carry over.
Reach
6,000+ connectors and plugins from the Claude Marketplace, plus its own browser for sites without one.
Cost
Included in Pro and Max. It uses your five-hour and weekly limits, and background checks count. New users get one full limit reset, usable until November 30, 2026.
Apps
iOS, Android and claude.ai. It's a separate app from Claude.
Gotchas that shape any demo
No Team or Enterprise plans yet. It's a personal Pro or Max product, so don't connect work accounts.
No Microsoft 365 (no Outlook or Exchange), and Slack support is limited. Demo on a throwaway Google account.
It checks on a timer. A planted email can sit unseen for a few hours, so seed the inbox the evening before and let the real morning brief run.
You can't watch or take over its browser, and it can't see your local files.
Per Every's review, scheduled runs can't reach custom connectors. Stick to the built-in Gmail, Calendar and Drive.
On October 23, Anthropic published "Claude Duet Safety, Privacy, and Security." It's the part of Duet worth holding every other agent to.
Control
What Anthropic says it does
Approval cards
Show the action, the destination and the full content, including every recipient. A separate check, built on Claude Code's auto-mode classifier, runs outside the model, so instructions hidden in an email can't approve their own action.
Always ask
Paying, deleting orders and creating accounts can never be set to "don't ask again."
Never, even with approval
Trading securities, moving money between accounts, and using someone else's account.
Purchases
One-time cards via Link from Stripe, locked to the merchant and the amount.
Sensitive info
A two-tier classifier: "never share" (credentials, full card and bank numbers, government IDs) and "needs consent" (health, finances).
Passwords
Typed into an in-app form, encrypted with a single-use key and decrypted only on the cloud-browser machine. The model never sees them. 1Password sign-in is "soon."
On the web
Duet's browser sends a distinct User-Agent from published IP addresses, and some large sites chose to block it before launch.
Testing
Pen-tested by an outside firm with source-code access, and in scope for Anthropic's HackerOne bug bounty.
For a demoMake the approval tap the climax. Open the card, show the full recipient list and content, edit one line, then approve. Say plainly that Duet can't send, pay, sign or delete without that tap.
Reviews and roundups from Duet's first week agree on a short list. The moment people tell others about is the agent noticing something they missed.
#
Use case
Why people like it
Demo value
1
The morning brief: "brief me on three things"
The flagship. The value is in what it leaves out.
Opener
2
The unprompted catch: a calendar clash, a reply you missed
The story every outlet repeats is a broker whose booking mistake Duet flagged without being asked.
Wow
3
"Who's waiting on me?"
Finds unanswered threads and drafts replies for your approval.
High
4
Living docs, dashboards and decks
Every's reviewer called building something and keeping it current the core of its value as a pro tool.
Closer
5
Chasing money: unpaid invoices, payroll prep for review
The small-business favorite, and the example Anthropic leads with.
Medium
6
Prep before a meeting
Simple and reliable, less impressive on stage.
Filler
7
Forecasts, models, prototype apps
Shows depth, but takes minutes and a lot of your plan's usage.
Pre-bake it
The eesel review (October 23) gave Duet 3.5 out of 5: strong prioritization, and "the approval line is drawn in the right place." Its complaints are the known limits: app checks only every few hours, no way to watch or take over its browser, no Microsoft 365, limited Slack, background work counts against your plan, and no Team or Enterprise plans.
Avoid on stage: personal errands and ticket booking (Anthropic says errands aren't its strength yet), polished-voice email drafting (Every rated its drafts below ChatGPT Dots), and anything that depends on a live check-in arriving on cue.
04 · Demo
The Store Support Copilot A demo for anyone who supports a fleet of retail stores.#
The persona: an IT or support person behind a chain of stores, buried in store tickets, vendor dispatches, opening checklists and managers asking for "a quick summary." Repeat issues are the problem everyone in that job feels, and spotting them is exactly the kind of noticing Duet is built for. Everything below is fictional.
The three planted catches, plus a bonus
Catch 1 · Silent SLA breach
The vendor moved the fix, and nobody said so
Store 9004 emails "URGENT, registers can't print." Three hours later, PrintCo's dispatch quietly pushes the visit to Monday. Saturday is the store's busiest day.
Duet shouldConnect the two emails, flag the broken 4-hour promise, and draft an escalation to PrintCo for approval.
Catch 2 · Repeat offender
The fifth ticket for the same LED panel
Store 9001: "Window LED flickering AGAIN." The tickets sheet holds exactly five window-LED tickets for that store in 90 days.
Duet shouldSpot that it's the fifth repeat and suggest replacing the panel rather than rebooting it again.
Catch 3 · Opening at risk
Six tablets, three days, one change freeze
Store 9008 opens in three days, its six tablets aren't enrolled in device management, the openings PM is asking, a change freeze starts Monday, and a regional VP is visiting.
Duet shouldJoin the email, the sheet and the calendar, and warn that the readiness check is tomorrow and the freeze cuts the time left.
Bonus · The clash
Two meetings and a newsletter
An 11:00 PrintCo vendor review overlaps an 11:30 call with the Store 9003 manager. There's also a newsletter in the inbox.
Duet shouldFlag the clash and skip the newsletter. Leaving it out is a feature worth pointing at.
What the brief should look like
7:00●●● 5G
DDuet7:00 AM
Three things need you today:
Store 9004 can't print receipts, and PrintCo moved the fix to Monday. That breaks their 4-hour SLA ahead of Saturday. I've drafted an escalation.
Store 9001's window LED is on its 5th ticket in 90 days. Worth asking LumenSign for a panel replacement?
Store 9008 opens Sunday, but its 6 tablets aren't enrolled, and the change freeze starts Monday. Readiness check is tomorrow at 2 PM.
Also: your 11:00 vendor review clashes with the 11:30 Store 9003 call.
EditReview escalation
Illustrative mock · not real Duet output
This is what a good brief should look like, not something Duet produced. Rehearse the demo, then screenshot the real brief for your backup slide.
Three things make it land in a room:
It connects sources. Catch 1 needs two emails; catch 3 needs an email, a sheet and the calendar. No single app would have flagged them.
It arrives with the next step drafted. The escalation is written; you only edit and approve.
It shows restraint. The newsletter isn't there, and nothing has been sent.
Paste this into Duet on day zero, before the first brief arrives. Anthropic says you can tell Duet what you want to hear about and how often, so be specific.
# Day 0 · the setup prompt
I run IT support for a chain of retail stores. Every weekday at 7am brief me on three things, most urgent first:
1. Store emails where a store is down or waiting on me, especially when a vendor has missed or moved a promised fix.
2. Stores reporting the same device issue repeatedly. Check the "tickets" sheet in my Drive and call out any store with 3+ tickets for the same device in 90 days.
3. Store openings in the next 7 days (the "stores" sheet, plus my calendar) where devices are not ready.
Also flag calendar clashes today. Skip newsletters and FYI emails where I'm only cc'd.
Draft replies and escalations, but never send without my approval.
# Channels to create (they share Duet's memory)#store-openings everything about Store 9008's opening
#vendor-escalations PrintCo, LumenSign, NetCare follow-ups
# Act 3 · the living doc (the closer)
Build me a one-page Google Doc for the regional director: which stores had the most tech issues this quarter, the top repeat issues, and what we're doing about each.
Use the tickets sheet. Keep it updated every Friday morning.
Everything is made up: stores 9001–9008, fictional vendors, an .example domain on every address, and the same data on every run. Build the dates relative to the day you run it, so "today" and "opens in 3 days" are always right.
Piece
What it holds
stores.csv
Eight fictional stores with opening dates and device counts, including Store 9008 opening in three days.
tickets.csv
About 40 past tickets across the stores, five of them the planted Store 9001 window-LED repeats within 90 days.
calendar.ics
About five events for "today," including the 11:00/11:30 clash and tomorrow's 2 PM readiness check.
Seven emails
Store 9004's urgent ticket, PrintCo's quiet reschedule three hours later, Store 9001's "AGAIN," the openings PM's question, the change-freeze notice, the VP visit, and a newsletter to ignore.
Seed script
Loads the emails straight into the demo inbox over IMAP with backdated times. Nothing is sent and nobody receives anything. Tag each message with a custom header so a reset script can remove them for the next rehearsal.
One-time setup (about ten minutes)
Create a throwaway Gmail account for the demo. Turn on 2-Step Verification and create an App Password for the seed script.
Join the Duet waitlist with the email on your Claude Pro or Max account. Once you're in, connect only the demo Gmail, Calendar and Drive.
Run the seed script the evening before, so the real morning brief picks the emails up.
Do the two clicks Google won't let a script do: import calendar.ics in Calendar settings, and upload the two CSVs to Drive as Google Sheets named "tickets" and "stores."
Paste the setup prompt from section 05 and create the two channels.
Want a coding agent to build the kit? Ask it for a seed script that writes the three files and seven .eml messages from the descriptions above, dates relative to today, plus up and reset commands. Keep every address on .example.
Not yet. There's no Microsoft 365 support and no Team or Enterprise plan, and company data would need IT and security sign-off anyway. That's why the demo is synthetic.
How is this different from a chatbot?
It's proactive and runs in the background on its own cloud computer. It messages you first and keeps working when your phone is off.
What stops it doing something dumb?
Approval cards on send, pay, sign and delete, checked outside the model; some actions can never be auto-approved, and trading or moving money is blocked outright. It's still a preview and can misjudge urgency, so check its priorities for the first few weeks.
A hostile email telling it to do things?
A real risk. The approval check runs outside the model so hidden instructions can't approve themselves, but the cards only protect you if people read them.
Who can get it, and when?
US Max users are coming off the waitlist first, then US Pro. No international date.
What does it cost?
Nothing extra on Pro or Max, but it uses the same limits, background checks included. Big jobs like models and decks use a lot.
Invitations go out in waves with no date promised. Run the same prompts in Claude (Cowork or chat with connectors) against the seed data, and say honestly that it's the "pull" version of the same idea.
The brief misses a catch
Ask directly in the thread: "Anything about Store 9004?" That still shows it reading across email and sheets.
Venue network fails
Have screenshots of the rehearsal brief and the doc ready.
Preview details change
Recheck eligibility, Microsoft 365 support and connector limits the day before.
Duet and n8n are a pair, not a choice
n8n is a deterministic pipeline you build and own, and it can run inside your network. Duet is a personal agent that decides what matters to you this morning. The companion n8n Ticket Triage Copilot demo tells the same repeat-ticket story from the pipeline side: n8n catches and routes every repeat on the way in, and Duet tells the person what needs them today.
Rule of thumbBuild the demo around one brief, three catches and one approval tap. Keep it personal, keep it synthetic.
Claude Duet is a separate app from Anthropic, in research preview since October 21, 2026, where Claude works as a proactive agent. It runs on Claude Opus 5.5 in its own cloud computer with its own browser, sends a short morning brief, checks your connected apps every few hours, and can use more than 6,000 connectors and plugins. It asks before it sends a message, pays, signs or deletes anything.
Who can get Claude Duet?
Duet is waitlist-only for US Claude Pro and Max subscribers, at no extra cost, and it uses the plan's usage limits, including its background checks. On October 28, 2026, Duet product lead Karan Ahuja said US Max users are coming off the waitlist first, with US Pro next. There's no date for international access, and it isn't on Team or Enterprise plans yet.
What do people actually use Claude Duet for?
Reviews published in its first week agree on a short list: the morning brief, the unprompted catch (a calendar clash or a reply you missed), finding who is waiting on you and drafting replies for approval, chasing unpaid invoices, and prep before a meeting.
Can Claude Duet spend money or send email on its own?
No. Approval cards show the action, the destination and the full content, including every recipient, and the approval check runs outside the model. Paying, deleting orders and creating accounts can never be set to "don't ask again." Trading securities, moving money between accounts and using someone else's account stay blocked even with approval. Purchases use one-time cards via Link from Stripe, locked to the merchant and amount.
Can I demo Claude Duet on my company's real tickets and Outlook?
Not today. Duet doesn't support Microsoft 365, has no Team or Enterprise plan, and company data would need IT and security sign-off. Run the demo on a throwaway Google account with synthetic stores, tickets and vendors instead.