Zugo AI: add AI to your app with window.zugoAI

Zugo AI puts a working AI runtime inside every project built in Zugo: any site, app, or game can call window.zugoAI from its own code while it runs, with no API keys and no backend. Fast mode costs about 20 calls per Zugo credit; smart mode, about 2. Starter credits are free, no card required.

Until now, a Zugo build could store data through Zugo Cloud or take payments through the Stripe connector. As of this release it can also talk back: answer a visitor's question in plain language, or play a character in your game.

What is Zugo AI?

Zugo AI is the built-in AI layer of every project created in Zugo: when your site, app, or game is built and published, it includes a global window.zugoAI object that sends a prompt to a language model and returns the response at runtime. It has two modes — fast, powered by Haiku, at roughly 20 calls per credit, and smart, powered by Sonnet, at roughly 2 calls per credit — and usage is metered in the same credits that pay for builds. There is nothing to install and no separate AI bill.

The usual blocker for adding AI to a small web app is the key. A browser app can't hold an API key safely, so even a toy chatbot normally needs its own server proxy plus a billing account with a model provider. window.zugoAI removes that layer: the call goes through Zugo, and no key ever appears in your client code.

You also don't have to write the call yourself. Describe the behavior in your build prompt — "add a helper that explains each chart in plain words" — and the agent wires window.zugoAI in for you, then verifies the result in the sandbox. Verified keeps its usual strict meaning here: the project actually loaded and rendered. And because editable code is free on Zugo, you can open the generated source and tune the call by hand whenever you want.

What's the difference between Zugo AI and the builder?

Zugo has always used AI at build time: you describe an idea, the agent generates the project — simple builds land in about a minute, complex ones take a few. Zugo AI is a different thing. It is the finished product calling a model while your users are inside it. The builder is the factory; window.zugoAI is an engine that ships inside the car.

That difference changes what's buildable. A static FAQ page answers the questions you predicted; an FAQ page with window.zugoAI answers the question the visitor actually typed. Game characters stop reading from a fixed dialogue table. Small utility apps — an explainer, a rewriter — become one-prompt builds instead of weekend API projects.

How much does a Zugo AI call cost?

Calls are priced in Zugo credits, the same currency that pays for builds. The two modes trade quality ceiling for volume:

For scale: Pro at $25/month includes 1,000 credits, which works out to roughly 20,000 fast calls or about 2,000 smart calls if you spent the whole allowance on Zugo AI. A game NPC that fields 200 player messages a day in fast mode burns about 10 credits a day. The free starter credits cover building an AI feature and testing it without a card; full plan details are on the pricing page.

A sane default: start every feature on fast, and move a call to smart only once you can see the quality difference inside your own app.

What does window.zugoAI look like in a real build?

Here is a build prompt, verbatim:

A cozy tavern mini-game: a pixel-art tavern with a keeper named Bram behind the bar. The player types anything into a chat box and Bram answers in character. Use Zugo AI in fast mode, keep replies to two short sentences.

The build came back verified in the usual minute or so. In the generated source, the agent had wired the call like this (excerpt):

async function askBram(playerInput) {
  const reply = await window.zugoAI.ask(playerInput, {
    mode: "fast",
    system: "You are Bram, a gruff but warm tavern keeper. Stay in character. Two short sentences max.",
  });
  addBubble("bram", reply);
}

And an exchange from the published app:

Player: any rumors going around?

Bram: Folk say the old mill light burns at midnight again. I'd leave that one be, friend.

Each of Bram's lines is one fast call, so a single credit buys about 20 of them. The personality lives entirely in that system line, and you change it the same way you change anything in Zugo: tell the builder "make Bram more superstitious" and rebuild. The sandbox re-verifies every iteration before you see it.

What can you build with it?

Four patterns that fit a one-prompt build:

A chat assistant for your product. A landing page that answers visitor questions from a system prompt you control ("answer only from the FAQ below"). Use fast for lookup-style answers; switch to smart if replies run long.

A text generator tool. The user pastes something in, your app hands back a rewrite — an email polisher, a caption writer. Here the output is the product, which makes it the clearest case for smart mode.

NPC dialogue in a game. Like Bram above. fast keeps conversation cheap enough that players can talk as long as they like, and one system line is enough for a consistent character.

A hint button. A quiz or puzzle that generates a nudge for the exact state the player is stuck in, instead of hard-coded hint text. One fast call per press.

For working ideas, open the showcase gallery — since July it supports upvotes and remixing, so you can take a published project apart and rebuild it your way.

What are the limits?

• Visitors spend your credits. Calls made by a published app draw from the owner's balance. If your app is public and popular, its AI feature has a real running cost — keep visitor-facing calls on fast and prompts short.
• Two modes, no custom models. Zugo AI is Haiku and Sonnet behind two switches. If you need a different provider or your own fine-tuned setup, export the code to GitHub and integrate it yourself — the code is yours.
• Verification checks the build, not the answers. The sandbox confirms the project loads and renders; it can't judge whether your assistant answers well. Quality lives in your system prompt, and a vague one produces vague replies.
• Model latency is real. A call takes as long as the model takes. Ask the builder for a typing indicator or a loading state so the wait reads as intentional.

What else shipped in July 2026?

Zugo AI is the headline of the July release, and it arrived with company: design directions before a site build (three previews to pick from), publication analytics with sources, devices, and pages, a security scan before publishing, automatic repair of runtime errors, and a referral program that adds 100 credits for each friend who upgrades to a paid plan. The full roundup is in the July 2026 update.

How do you try it?

Describe an app with one AI feature in a sentence or two and build it. The first working version takes a minute or two, gets verified in the sandbox, and publishes in one click to yourname.zugo.run. Starter credits are free with no card, so a fast-mode prototype costs nothing to test. If this is your first build, start with the walkthrough in how to build a working app with AI, then give your prompt one extra sentence: what the AI inside your app should do.