What is Caffeine.ai?
With Caffeine.ai, the Dfinity Foundation has released a platform that takes a different approach from Copilot, Cursor or Claude. Instead of helping developers write code, the AI takes over the coding entirely. You describe in natural language what you want, and an ensemble of AI models writes and deploys the application.
What's interesting: the apps run on the decentralized Internet Computer Protocol (ICP), which the Dfinity Foundation itself launched in 2021, and the purpose-built programming language Motoko mathematically guarantees that updates cannot lose data. At a hackathon, non-technical people have already built working apps, from a dentist with a patient records tool to a knitter with a hobby learning app.
A working example of Dfinity's vision is OpenChat, a messaging app that runs entirely on the Internet Computer and is governed by a DAO with tens of thousands of participants.

According to VentureBeat, the whole thing is backed by heavyweight investors: more than 100 million dollars from Andreessen Horowitz and Polychain Capital went into the project. Caffeine is operated by Caffeine Labs AG, based in Zurich.
Cool: clone & change apps from an app store
"App stores today obviously operate on gatekeeping – that's going to erode", said Dfinity's Chief Business Officer Pierre Samaties at the World Computer Summit. Instead of buying apps, users can clone them and adapt them for their own purposes. A fundamentally different model from the Apple App Store or Google Play. This is what it looks like in the marketplace:

My test: a domain monitoring tool in 10 minutes
This is how Caffeine.ai greets its users, with a minimalist chat interface and the prompt "Chat to create!":

My spontaneous idea: a domain monitoring tool that shows me whether domains are about to expire or have been dropped. Anyone who owns dozens of domains like I do knows the problem: you quickly lose track.
My first prompt
A single simple sentence is enough: "create an app to monitor if a list of domains - are they dropped or expired"

Caffeine briefly asks for details (check daily? dashboard with colors?) and then gets going.
The SPEC.md: Caffeine plans before coding
Before Caffeine writes a single line of code, it creates a detailed specification:


The chat: iterating in conversation
After the specification, things get serious. Caffeine shows its progress: "Initiating app creation", "Creating your app code from scratch", "Building your app and pushing to the network". A few minutes later: "Version 1 is up and ready for testing."
I notice a few problems: WHOIS data is missing, and the list doesn't show all domains. Some quick feedback in the chat, and Caffeine iterates:

A look under the hood: Motoko code
If you're curious, you can look at the generated code. Caffeine writes Motoko, the programming language of the Internet Computer. Here you can see the DomainEntry structure with status, WHOIS data and HTTP checks:

The result: not bad for 10 minutes
The finished app offers a grid and a list view, status indicators (Active/Pending/Expired) and a clear dashboard:


The WHOIS data still shows as "Unknown", so another iteration would be needed here. But as a proof of concept for 10 minutes of work? Impressive.
For completeness, here is the landing page in the mobile view, and of course there is a light/dark mode too:

Pleasantly surprised: login with Internet Identity
What really impressed me: logging in with Internet Identity (ICP's decentralized auth system) works flawlessly. Three steps and I'm in:
1. Choose a login method: Google, Apple, Microsoft or passkey. I choose passkey:

2. Name the identity: give it a name once, done. No email, no password:

3. Confirm and you're logged in: that's it. The app recognizes my identity right away:

Brilliant that auth is this easy for an app that was just created.
Credits & costs
Without signing up you get 5 credits; that was enough to iterate on the app twice. The credits reset daily. After registering you get another 110 credits for free.

Conclusion
Caffeine.ai looks exciting so far. For quick prototypes it already works surprisingly well. The prototype was fairly simple, and what I showed here also works very well with the various IDEs, but I like the combination of natural language, automatic specification and direct deployment.