AI teammates like OpenAI’s Dots, Meta’s Muse, and Grok’s Bot are, when you get right down to it, just AI with its own computer.
I built this whole setup from scratch. It took just over two hours, and I’ve condensed the entire process into this diagram.
It can book flights and write emails to suppliers to negotiate prices. When it gets stuck because it doesn’t know something, you can take over its desktop and fill it in for it. And things you tell it in a private chat? It can remember them in group chats, too.
Here’s how to read the diagram.
The top row has three architecture diagrams. Start on the left for an overview of the system and who’s talking to whom.
The middle one shows how a message travels from the chat box to the AI and back. The one on the right shows how you can view and take over its desktop—how you see its screen and step in.
The 12 cards below show the build sequence, from left to right, top to bottom.
There’s just one main thread through the whole diagram: first give the AI a library of ready-made primitive components, skills, and existing examples; then let it build features and commit when it’s done. Starting from an empty project, build it in order: sign-in with Clerk, database with Neon, the chat interface, then add a refresh-proof background worker and a cloud computer for each bot with Daytona. Finally, add group-chat memory, billing, and error monitoring with Sentry. It’s not that hard.
If you want to focus on just one part, look at the green box below the architecture diagrams: why can’t the AI run in the website backend? Why does it need a separate, stateless worker?
Refreshes, crashes, and deployments can all interrupt long-running tasks. That’s a design constraint you can’t get around. To avoid some common pitfalls, scan the red labels: free-tier sandboxes have no internet access; the AI may switch branches on its own; and a deployment can fail because a Next.js template included in a skill gets mistaken for a second app.
The text is small, so zoom in. It’s really useful. $NVDAB
I built this whole setup from scratch. It took just over two hours, and I’ve condensed the entire process into this diagram.
It can book flights and write emails to suppliers to negotiate prices. When it gets stuck because it doesn’t know something, you can take over its desktop and fill it in for it. And things you tell it in a private chat? It can remember them in group chats, too.
Here’s how to read the diagram.
The top row has three architecture diagrams. Start on the left for an overview of the system and who’s talking to whom.
The middle one shows how a message travels from the chat box to the AI and back. The one on the right shows how you can view and take over its desktop—how you see its screen and step in.
The 12 cards below show the build sequence, from left to right, top to bottom.
There’s just one main thread through the whole diagram: first give the AI a library of ready-made primitive components, skills, and existing examples; then let it build features and commit when it’s done. Starting from an empty project, build it in order: sign-in with Clerk, database with Neon, the chat interface, then add a refresh-proof background worker and a cloud computer for each bot with Daytona. Finally, add group-chat memory, billing, and error monitoring with Sentry. It’s not that hard.
If you want to focus on just one part, look at the green box below the architecture diagrams: why can’t the AI run in the website backend? Why does it need a separate, stateless worker?
Refreshes, crashes, and deployments can all interrupt long-running tasks. That’s a design constraint you can’t get around. To avoid some common pitfalls, scan the red labels: free-tier sandboxes have no internet access; the AI may switch branches on its own; and a deployment can fail because a Next.js template included in a skill gets mistaken for a second app.
The text is small, so zoom in. It’s really useful. $NVDAB