← All projects
Vendor-Agnostic Tooling active

Niu Prototyper

Visual interactive prototyping for live web UIs: wraps any static build or dev server in an inspectable canvas where I pin element-anchored comments and iterate in place with any CLI coding agent.

This came from micro-design feedback in user research sessions. In a rapid prototyping session, “move the second subtitle in the right-hand card” is a design I need to evaluate in real-time with a user.

Prototyper reviews live websites with a coding agent. It wraps a local build or dev server in a browser canvas, where I can select an element, pin a comment, and review changes in place.

How it works

Running web UI Dev server or static build wrapped with an injected DOM bridge PREVIEW ORIGIN Inspect & pin a comment Click any element; bridge captures selector path, content hash, visible text, and HTML snapshot across origin-checked postMessage SHELL ORIGIN Claim batch & read live DOM Claude, Codex, or Antigravity claims the queue in the tmux drawer and captures the live rendered DOM around the target element CLI AGENT Write a page overlay Injects scoped HTML/CSS over the target mount to test a layout without touching source files OVERLAY or Edit project source Updates component or stylesheet in repo; dev server hot-reloads or static build triggers reload SOURCE Preview updates in place Agent replies addressed; I verify the live UI and resolve the thread LIVE REVIEW

Token Economy

The agent never needs the whole codebase to act on a comment. Each comment carries the element’s selector path, a 160-character text excerpt, its outerHTML capped at 2,000 characters, and a content hash. The agent starts from that one element and searches the source, eliminating the need to read the whole site.

Point at the thing

Inspect mode captures the selected element’s location in the page, its text, and a snapshot of its HTML. The comment carries that evidence so the agent can find the intended target in the source.

Feedback goes into a queue. An agent claims the batch, inspects the target, changes the component or stylesheet, and reloads the preview for review.

Try a change

For a quick layout experiment, the agent can apply an HTML/CSS overlay. For a lasting change, it edits the project source. That lets me try a variation before deciding to build it into a component.

Built on Canvas

The terminal drawer comes from Document Canvas, so I can use the same CLI agents for prose and interface work.

The preview and the review shell run on separate local origins. Their messages are checked so the target application’s scripts cannot read the shell’s session token.

Stack

PartTool
Shell, dev proxy, and live reloadNode.js (HTTP server, reverse proxy, Server-Sent Events)
Preview isolation and DOM captureDual loopback origins with origin-checked postMessage bridge
Terminal drawer and agent session engineShared document-canvas engine (Python, tmux, node-pty, ws, xterm.js)
VerificationNode test runner and Playwright end-to-end smoke suite