An AI-first web framework.

Invalid programs are hard to express. Valid programs are cheap to verify.

Give your coding agent a structure it can check: typed views, explicit data and contracts for every behaviour.

Start building Understand Hozu

ほぞ / A joint that fits.

Your first 30 seconds

Create an app, add a feature, check your work. Node 22.18 or newer.

Use --agent agents for Codex, Cursor or Copilot; --agent claude for Claude Code.

bash

npm create hozu@latest my-app -- --agent claude
cd my-app
npm install
npx hozu add feature tasks --page /tasks
npx hozu check

Structure you can verify.

The IR is the source of truth

Typed builders describe your app as data. The validator, compiler and runtime share one representation.

Contracts for every behaviour

Given a state, when an event happens, expect a result. Every machine transition has a contract.

Rendering is derived

Declare who can see data and how fresh it must be. Hozu derives the render plan; only machine-bound views hydrate.

Diagnostics that lead to a fix

Each diagnostic carries a location, a cause and a suggested fix. Read it yourself or pass the JSON to your agent.

Hozu is not a separate SPA mode or a place for arbitrary effects in views. It makes behaviour explicit, and asks you to pay for that structure.

Measured, with the rough edges included.

A notes app with accounts, built from a spec and then changed by an agent, checked by a hidden acceptance test: per-user isolation, double submit, forms without JavaScript, HttpOnly sessions, and a regression pass after the change.

Trials 0012 and 0016: the same task, prompts and model
	Hozu 0.5	Nuxt
Checks passed	180 / 180 (5 runs, incl. Codex)	67 / 72 (2 runs)
Agent cost to build	1.38×	1×
Agent cost to change	1.45×	1×

One Nuxt change silently broke three working features; no Hozu run did. The price is extra tokens, mostly spent reading the guide of a framework the model has not seen. Small samples, not failure rates.