How to Set Up Duet: A Founder Walkthrough for Your AI Workforce

This guide is the exact walkthrough David, Duet's founder, gives to new customers on their first call. It is not a feature tour. It is the order of operations that turns a new workspace into a working AI coworker — what to do first, what to skip, and the mindset shifts that separate the people who get 10x value from the people who treat Duet like a fancier ChatGPT.

If you just signed up, read this once before you do anything else. If you're an implementation partner setting up Duet for a client, this is the playbook.

1. Memory matters more than the model

Every Duet workspace ships with an onboarding chat. By default it scrapes your website and writes a first pass of memories about your company. That's the floor, not the ceiling.

The single biggest predictor of how useful your Duet workspace becomes is what happens next: what you tell it about yourself in the first hour.

Think of it like this. If you transplanted every memory you have into another person's brain, that person would behave a lot like you — same opinions, same instincts, same shortcuts. The model is the brain. Your memory is what makes Duet act like your coworker instead of a generic assistant.

The onboarding chat asks proactive questions. Answer them properly. Lazy answers in the first hour cost you for the next year.

If you came from Aomni, this is the same idea as "model training" — except Duet does it implicitly through conversation and explicitly through the memory file, and the word "memory" is more accurate than "training." You're not retraining the model. You're giving it context.

2. Garden your knowledge base

The Files tab in your workspace is a real file system on a real cloud server. You can upload PDFs, drag in folders, paste URLs (Duet will scrape them into markdown automatically), and Duet can read everything in there.

The way to think about it: set it up the way you would organize a Google Drive for a brand new employee.

/files
  /company
    one-pager.pdf
    pitch-deck.pdf
    style-guide.md
  /customers
    target-icp.md
    case-studies/
  /products
    pricing.md
    roadmap.md
  /sales
    discovery-questions.md
    objection-handling.md

This is what David calls "gardening." You don't have to do it constantly. A quick sanity check every couple of weeks is enough — move anything that obviously belongs somewhere else, prune anything stale.

A well-organized memory file links to sub-memories, which link to other files. The agent navigates your file system the same way you do. Better structure = smarter coworker.

3. Turn repeatable work into skills

A skill is a memorized, repeatable workflow. It's the thing that turns one good run into a reusable capability for the whole workspace.

Here's how skills actually get created in practice. Someone on the team does a great piece of work in a chat — say, an SEO audit on a blog post. They ask all the right questions: are there internal links, what's the keyword cannibalization risk, what's the AEO posture, how does this rank on DataForSEO. The chat ends with a really good output.

Instead of letting that line of expertise live in one chat that nobody reads again, you say:

"Turn this SEO audit workflow into a skill we can all use."

Duet writes the skill — sometimes just a prompt, sometimes a prompt plus actual scripts that hit data sources. Now anyone in the workspace types /seo-audit and gets the same expert workflow, with the expert's domain knowledge baked in. The next time someone writes a blog post, they don't have to find the SEO person. The skill is the SEO person, for that one task.

This is the part that doesn't exist in single-player AI chat. In ChatGPT, you talk to the AI alone. In Duet, you watch other people work, they watch you, and the best work gets captured as shared muscle memory.

A skill is personal software for your company. Your SEO skill should look nothing like another company's — different keywords, different priorities, different data sources. That's the point.

4. Schedule recurring work with Relays

Most agents handle recurring work with cron jobs that run somewhere you can't see. Duet's approach is the opposite. Scheduled work lives inside a channel as a long-lived AI conversation that wakes itself up.

These are called Relays. (The old "cron jobs" tab is being deprecated.)

Here is what it looks like in practice. You create a channel called #scheduled-tasks. You give the AI an instruction:

"Every hour, check our inbox. Every day at 9am, post the growth report. Every Monday, run the weekly SEO check."

Duet sets up the Relay. From then on, you see the messages in the channel: "Reminder to continue scheduled task in 14 minutes." "Inbox checked — 2 unread, posted summary to #general." "Daily growth report posted to #growth."

The big win over hidden cron jobs:

• You can see what it's doing. Not "hopefully it ran."
• You can tweak it in conversation. "Actually, also check Reddit while you're at it." No re-config UI.
• You can pause or cancel it inline. It's just messages.
• You can change the model per Relay the same way you change the model in any other channel — with the model selector.

5. Connect any tool with an API

Duet ships with around 500 preconfigured integrations — Google Calendar, Slack, Notion, Linear, Stripe, Gmail, Telegram, you name it. Browse them in the Integrations tab and connect with one click.

What most people miss: Duet can connect to anything with an API, not just the 500 preconfigured ones.

The way it works: Duet is, at its core, a coding agent. Every AI agent is. Even when you ask it to do "research" or "sales prospecting," it's actually writing and running code under the hood. So if you want to connect a random fifteen-year-old billing system that nobody else integrates with, you just ask:

"Does Sentry have an API? Connect to it. Here's our API key."

Duet looks up the docs, writes the integration, stores your token in an environment file (the same way a human developer would), and calls the API. Five minutes later you're querying that system inside Duet. No MCP server required — MCP only works if the other party opens an endpoint. Code works always.

6. Ship internal and external apps

Because Duet has its own server and runs a coding agent, it can build apps. Real ones.

There are two flavors and you'll probably use both:

Internal apps. A dashboard, an admin tool, an "internal use cases" page that links to your best workflows. Duet builds them, hosts them on your server, and you can share the URL with your team. We use this constantly — every cool tab in our own Duet workspace ends up linked from one internal app we maintain.

External apps. A landing page, a marketing site, a customer portal. Same flow. You ask, Duet ships, the app gets a .duet.so URL by default, and you can wire your own custom domain (like consigliere.com) by setting up the DNS record. We currently host customer-facing marketing sites for several companies entirely inside their Duet workspace.

Hosting, custom domains, SSL — all included. No "deploy" button to find. The app is just there once Duet writes it.

7. Security and your own server

The single most important thing to understand about Duet's security model: every workspace is its own server.

Not a row in someone else's database. A real, isolated, single-tenant virtual machine. 100 GB of storage. 16 GB of RAM. 8 cores. Files on it are yours, not shared with other customers.

This isn't a marketing claim, it's an architectural requirement. The reason Duet can update its own skills, write new code, install new tools, and remember everything across sessions is because it owns its own computer. Multi-tenant agent products literally cannot work the way Duet works. They have to share infrastructure, which means they have to lock down what the agent is allowed to touch.

Practically:

• API tokens are stored in environment files, the same way a developer would store them.
• The agent is blocked from reading its own env file at runtime — env vars are injected into subprocesses, not handed to the LLM.
• You can browse, edit, and download every file on your server. You can pull a terminal. You can back up the whole machine.
• The underlying infrastructure runs on standard cloud providers (AWS, Google Cloud, etc.) via a sandbox provider — same baseline as any B2B SaaS.

If you're at a company with compliance requirements, this single-tenant model is usually easier to defend than a typical multi-tenant agent platform, not harder.

8. Billing without surprises

Duet is usage-based. Two things to know.

Why usage-based at all. The price of running an AI coworker is closer to the price of a human employee than a SaaS seat. Our top customer spends about $4,500/month. Other customers spend $5/month — they log in three times and ask a question. There is no per-seat number that makes sense across that range. So we just pass through token cost. We negotiate bulk discounts with the model providers; we pocket the discount. You pay the same as if you went to OpenAI or Anthropic directly.

How to avoid sticker shock. Three settings, all in your billing page:

The dashboard breaks usage down per member, per day, with per-message tags. You can always see what cost what.

9. Pick the right model

Duet supports the frontier models — Claude Opus, GPT, Gemini, and the rest. The model selector is in every channel.

Default to Opus when you're getting started. Especially in the setup phase. Better memory, better instruction following, better at the "is this skill worth creating" decisions. Yes, it's more expensive. The amount of time it saves in the first month dwarfs the token cost.

Once your memory and skills are tuned, try cheaper models for the workloads that don't need Opus — Q&A over your knowledge base, drafting a quick reply, summarizing a thread. You can set the default model per workspace; new channels inherit it.

We've tested automatic model routing in the past. It didn't work well. The models have very different personalities, and an auto-router produces a "schizophrenic" experience — sometimes great, sometimes bafflingly off, no way to know why. For now, choose the model explicitly.

10. External channels: email and Telegram

Duet can map external messaging systems directly into channels.

Email. Connect an inbox in the Integrations tab. Every email that lands in that inbox becomes a message in a dedicated channel. You set a "first submit" prompt — what should the agent do when a new email arrives? Common patterns:

• "Summarize the email and recommend a next action. Don't reply automatically."
• "Reply to all customer support emails using our brand voice. Escalate anything refund-related to me."
• "File the email under the right contact's thread and add a one-line summary to memory."

Then you talk to Duet in the channel like any other chat. "Reply to him and say X." "What's the status of this thread?" "Mark all of these as resolved."

Telegram. Simpler. Duet connects to a Telegram bot, and anyone who messages that bot is treated as an internal user with access to your workspace. No prompt configuration — it's a direct bridge.

Email is more powerful because email talks to both external customers and internal team. Telegram is for internal-only.

11. White labeling and renaming the AI

Duet is much more customizable than people expect. You can:

• Rename your workspace (in the workspace settings, formerly called "server settings").
• Upload your own logo to replace the Duet logo.
• Rename the AI itself. Mentions update automatically — if you rename the AI to "Consigliere," @duet becomes @consigliere everywhere in your workspace, and the AI knows itself by that name.
• Use a custom domain for any app Duet builds.
• Set up DNS so your customers see your brand, not ours.

The only thing we don't currently support fully white-labeling is the root Duet app URL. The pragmatic workaround: build your customer-facing experience as an app inside Duet, point your own domain at it, and the customer never sees the word "Duet."

12. The mindset shift

If you take one thing away from this guide, take this:

Using AI well is a little like gambling.

You're betting some tokens — call it $1, or $10, or $100 — that the AI will produce something that returns multiples of that value. Sometimes it doesn't. Sometimes the run goes sideways and you've spent $20 on nothing. But over a month, if you know roughly what you want, the expected value is heavily positive. Our top users spend hundreds of dollars in tokens a month and get tens of thousands of dollars of work back.

This mindset only works if you know what you want. Vague prompts on expensive models with no memory in place will burn money fast. Specific outcomes, with memory and skills in place, will pay for themselves in week one.

The corollary: don't try to learn AI by exploring randomly. Pick one boring, valuable, repeatable workflow in your business — answering customer support emails, doing weekly SEO audits, updating your website, drafting outbound sequences, building an internal dashboard — and make Duet good at that one thing. Then add the next. People who try to "use AI for everything" on day one usually use it for nothing by week four.

13. FAQ