Huzzah

// Overview

Huzzah inverts the usual arrangement between a prompt and a program. In a chat-driven workflow the source file is the artefact and the prompts that produced it evaporate; in Huzzah the pseudocode is the artefact that lives on disk in `.hz` files, and the generated implementation is what gets reconciled against it. You edit intent, and the editor keeps a real implementation synchronized with what you wrote.

Because the pseudocode persists, each change is a diff of intent rather than a fresh description of the whole program: an edit sends only the pseudocode diff as the prompt, and the file that remains afterwards reads as a record of what the program is supposed to do. Accepted AI-generated JavaScript runs locally in a Web Worker, which the author is careful to describe as experimental containment rather than a sandbox for hostile code — the README's own advice is not to paste secrets or private source into it.

It is a browser app you run locally with `npm run dev`, and it is vendor-neutral: model access goes through Pi's provider, credential and model configuration, so it works with an API key for Anthropic, OpenAI, Google, Azure OpenAI, Amazon Bedrock and others, with a subscription login through Pi, or with a local model served by Ollama, LM Studio or vLLM. The project describes itself as experimental, and the README suggests small self-contained programs — a discount-applying shopping cart, a rate limiter, a turn-based combat system — as the way to get a feel for it.

// What it does

• Pseudocode is the persistent artefact: `.hz` files stay on disk as a record of intent, with the implementation derived from them
• Each edit sends only the pseudocode diff as the prompt, instead of re-describing the whole program
• Vendor-neutral model access through Pi — API key, subscription login, or a local model via Ollama, LM Studio or vLLM
• Accepted AI-generated JavaScript executes in a Web Worker (experimental containment, not a hostile-code sandbox)
• Semantic highlighting and reconciliation run at Pi's `low` thinking level, keeping the edit loop cheap
• Runs locally as a browser app on a dev server — no account, no hosted backend

// Getting started

You need Node.js 22.19 or newer and access to a model provider supported by Pi. Huzzah reuses Pi's provider, credential and model configuration rather than defining its own.

git clone https://github.com/danielvaughn/hz.git
cd hz
npm install

export ANTHROPIC_API_KEY=sk-ant-...

npx --workspace=poc pi
# then: /login

cat ~/.pi/agent/settings.json
# { "defaultProvider": "anthropic", "defaultModel": "claude-sonnet-4-5" }
npm run dev

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Keep a durable, reviewable statement of intent next to generated code instead of losing it with the chat history
• Iterate on a small self-contained program — a rate limiter, a scheduler, a scoring rule — by editing the rules rather than the implementation
• Try a spec-first workflow against your own provider, including a local model, without signing up for a hosted product
• Study an alternative to chat-driven coding: reconciliation from a persistent spec rather than turn-by-turn prompting

// How Huzzah compares

Huzzah alongside other open-source ide copilots tools AI/TLDR tracks, ranked by GitHub stars.