Claude Duet: What It Does and What to Build Yourself

By Lazar Milicevic · Published October 23, 2026 · 11 min read

You run a small business and your inbox, Slack, invoices, and calendar all want attention at once. Anthropic's answer, announced October 21, 2026, is Claude Duet: an agent that watches your connected apps and tells you what matters. Before you join the waitlist or plan around it, here's what it actually does, what it costs you, where it's likely to fall short, and when you should build your own agent instead.

What Claude Duet is, in plain terms

Claude Duet is a proactive agent app from Anthropic, released as a research preview on October 21, 2026. It runs as a standalone app on iOS and Android and also on claude.ai. Access is by waitlist for Pro and Max subscribers in the US. Team and Enterprise plans are not included in the preview yet.

"Proactive" is the part that separates it from a chat window. You don't open Duet and ask a question. It connects to your accounts, checks them on a schedule, sends a short morning brief covering the most important items across your connected apps, and then messages you during the day only when something needs attention.

According to TechCrunch, Duet is built on Claude Opus 5.5 and can connect to the Claude Marketplace. Anthropic positions it for professionals and small business owners who have more on their plates than they can track, which is exactly the audience that tends to read this blog.

One framing correction: some coverage calls Duet Anthropic's "first consumer-facing AI agent." Anthropic's own positioning is professionals and small business owners, so I'd treat it as a work tool, not a consumer gadget.

Sources: Introducing Claude Duet, TechCrunch.

The "6,000 connectors" number needs an asterisk

TechCrunch reports Duet can connect to over 6,000 connectors and plug-ins in the Claude Marketplace. That's a Marketplace count, not a list of what Duet does well on day one.

When TECHi went looking, Anthropic named only about 13 integrations by name, including Gmail, Slack, QuickBooks, Stripe, and Gusto. That's a very different planning number than 6,000.

The practical reading:

• Named integrations (the dozen or so Anthropic lists) are the ones to evaluate. Check whether your actual stack is on that list.
• The rest of the Marketplace may be reachable in principle, but "a connector exists" is not the same as "Duet uses it reliably to triage your day."
• Any tool you depend on that isn't on the named list is a gap you should test, not assume.

Before you spend time on the waitlist, write down the five systems where your real work lives. If three of them are named integrations, Duet is worth a trial. If none are, you're evaluating a roadmap, not a product.

How it behaves: polling, briefs, and limits

Duet checks connected accounts every few hours rather than continuously. The reason is plan limits: nonstop checking would burn through your usage allowance much faster. Two consequences worth understanding:

1. It is not real-time. If a client emails at 10:02 and Duet last checked at 9:30, you may not hear about it until the next sweep. For genuinely time-critical alerts (a payment failure, a downtime page), keep your existing push notifications or a purpose-built automation.
2. Background checks count toward your limits. Duet is included in Pro and Max with the same usage limits as the rest of Claude, and Anthropic doesn't list a separate Duet price during the preview. Heavier work, such as building a financial model or a deck, takes more model turns and uses more of the allowance.

So the real cost of Duet is not a line item. It's the usage headroom you lose for everything else you do in Claude. If you already hit limits on Pro, adding an agent that polls in the background and takes on large tasks will make that worse.

On price: check the current pricing page on claude.com for Pro and Max plan costs before you decide. Anthropic's Duet announcement lists no separate Duet price.

What it can do on your behalf, and what it refuses

Agents get risky when they can act, so the permission model matters more than the demo. Based on reporting of Anthropic's announcement:

Duet can also make business purchases through Stripe's Link or Shopify's Shop Pay, with approval steps. Link uses virtual-card handling, which limits exposure of your real card details.

A few caveats on this table. The safety details (the per-message classifier, the safety system prompt, the no-credentials design) are self-reported by Anthropic, and I'm not aware of independent assessment yet. Fuller safety details were promised in the days after launch. Treat the approval gates as a reasonable design, not as proof of safety.

What I like about the design is that the approval step is the default for every irreversible action. That's the right instinct. What I'd still do: start with read-only connections and add write permissions one at a time, so you learn how it behaves before it can send anything in your name.

Where Duet will likely struggle

Anthropic itself says Duet can still misjudge what is urgent and that it will keep tuning it based on feedback from early users. That's an unusually candid admission, and it points at the core difficulty of any triage agent: urgency is contextual.

Think about what "important" means in your business. A message from a vendor you've never heard of might be spam or might be your supplier warning about a delay. A short reply from a client might be a polite thank-you or the start of a cancellation. Duet doesn't know your relationships the way you do, at least not at first.

Expect failure in two directions:

• False negatives: something important sits unflagged until the next check, or never surfaces.
• False positives: the morning brief fills up with noise, and you stop reading it. A brief you ignore is worse than no brief, because it gives you false comfort.

A simple way to test this during the preview: for two weeks, keep your normal routine and compare. Each evening, write down the three things that actually mattered that day. Check whether Duet's brief surfaced them. If it catches two of three most days, it's useful. If it misses the same category every time, you've learned where it can't help yet.

Duet vs. building your own agent

This is the decision most small teams actually face. Duet is a finished product with a fixed behavior. A custom agent, built on the Claude API with your own integrations, is something you shape. Here's how I'd think about the tradeoff.

A key point: as far as Anthropic's published pages go, Duet is an app, and I haven't seen confirmation of an API for it. One source says there is none, but that is the author's own observation, not something on Anthropic's list. And Team and Enterprise access is described only as "not included yet," with no date given. If you need Duet-like behavior for several people, or inside your own systems, you shouldn't wait on it.

Adopt Duet when: you're a solo operator or tiny team, your tools are on the named integration list, you value zero setup, and you can live with occasional misjudged urgency.

Build your own when: your triage rules are specific to your business (e.g., "flag any invoice over a threshold from a new vendor"), you need audit logs, you need alerts faster than a few hours, or you need the agent to run for a team.

What a custom version costs to run

If you build instead, you pay API rates. Rates for Claude Opus 5.5 are published in Anthropic's official docs, along with prompt-caching rates, a Batch API discount, context window and output limits, and options such as fast mode and US-only inference. Check the Opus 5.5 docs for the current figures and launch details before you budget, since pricing can change and secondary coverage sometimes frames the comparison to earlier models differently than Anthropic's own pages do.

Here is an illustrative example with round numbers, not a benchmark. Suppose a triage agent processes a batch of messages each sweep, and each sweep sends 50,000 input tokens and gets back 2,000 output tokens. Fill in the per-million-token rates from the pricing page:

# Illustrative only: round numbers, not a measurement
input_tokens  = 50_000
output_tokens = 2_000

# Replace these with the current per-million-token rates from the docs
INPUT_RATE_PER_M  = ...
OUTPUT_RATE_PER_M = ...

input_cost  = input_tokens  / 1_000_000 * INPUT_RATE_PER_M
output_cost = output_tokens / 1_000_000 * OUTPUT_RATE_PER_M

per_sweep = input_cost + output_cost
print(f"Per sweep: ${per_sweep:.2f}")
print(f"Six sweeps/day: ${per_sweep * 6:.2f}")

Prompt caching changes this a lot. If most of each sweep is a stable system prompt plus standing instructions, cache reads are far cheaper than fresh input. Batch processing helps further for anything that isn't urgent, since the Batch API is discounted.

The point of the math isn't the exact number. It's that a scoped, cached triage agent can be cheap, and you control the cadence. Duet's cost, by contrast, is hidden inside your plan limits.

A minimal triage agent you can run yourself

If you want to prototype the core idea before deciding, here's a small skeleton using the Anthropic Python SDK. It takes a list of messages and returns a ranked brief. Replace the model identifier with the one from the official Opus 5.5 docs.

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

SYSTEM = """You triage a small-business owner's inbox.
Rank items by urgency. Flag only items that need a decision today.
Rules:
- Any message mentioning a failed payment or cancellation: urgent.
- Newsletters and receipts: never urgent.
Return JSON: [{"id": str, "urgency": "high|med|low", "reason": str}]"""

def triage(messages: list[dict]) -> str:
    body = "\n\n".join(
        f"[{m['id']}] From: {m['from']}\nSubject: {m['subject']}\n{m['snippet']}"
        for m in messages
    )
    resp = client.messages.create(
        model="YOUR_MODEL_ID",  # replace with the identifier from the official docs
        max_tokens=1024,
        system=SYSTEM,
        messages=[{"role": "user", "content": body}],
    )
    return resp.content[0].text

This does one thing: it ranks. It sends nothing, deletes nothing, and pays nothing. That's deliberate. The first version of any agent should be read-only and should output something you can review. Once the ranking matches your own judgment for a couple of weeks, you can wire it to a schedule and a delivery channel.

Notice the rules in the system prompt. That's the advantage over Duet: "a failed payment is always urgent" is a rule you write, not a behavior you hope the vendor tunes toward.

A decision checklist before you commit

Run through these before joining the waitlist or starting a build:

1. List your five core tools. How many are named Duet integrations?
2. Check your plan. Are you on Pro or Max in the US? If you're on Team, Enterprise, or outside the US, you can't use the preview today.
3. Check your headroom. Do you already hit usage limits? Background polling and large tasks will compete with your normal use.
4. Define "urgent" in writing. If you can't state your triage rules in ten lines, no agent will guess them well.
5. Decide the blast radius. Which actions are you comfortable letting any agent take, even with approval? Start read-only.
6. Set a two-week test. Compare the agent's brief with what actually mattered. Decide on evidence.

Where this leaves a small team

I'm a senior engineer who builds AI automations for small teams, so the Duet-versus-build question is one I think about regularly. The pattern I see: an off-the-shelf agent is the right call for standard stacks and single users, and a custom build earns its cost when your rules, audit needs, or tool list don't fit a fixed product.

If you're weighing Duet against building something scoped to your workflow, get in touch through BizFlowAI. I'd rather you pick the cheaper option that works than the more elaborate one.

The bottom line

Claude Duet is a real step toward agents that work in the background for professionals and small business owners, with sensible approval gates and a candid admission that urgency judgments are still being tuned. It's also a research preview limited to US Pro and Max subscribers, bound by shared usage limits, polling every few hours, and dependent on integrations that are far fewer than the headline Marketplace count suggests.

Join the waitlist if your stack matches the named integrations, and run a disciplined two-week comparison. If your needs are specific, your team is larger than one, or you need faster alerts, a scoped custom agent on the API is cheap to run and fully yours to control. Either way, start read-only, keep humans on every irreversible action, and judge the agent by what it actually catches, not by the demo.

Work with BizFlowAI

If you'd rather have this built for you, that's what we do: production AI automation for solo founders and small teams — agents, integrations, and document pipelines that actually ship.

Request a free discovery call — 30 minutes, we map the highest-ROI automation in your workflow. No pitch deck, just engineering.

More guides like this on the BizFlowAI blog.