A desktop browser for designers and developers whose browsing should leave something behind.
Extract mode — hover the live page, walk the DOM with the arrow keys, click to keep everything about a region.
You find a pricing section you love. You screenshot it. Three months later you have a folder of
4,000 PNGs named Screenshot 2026-04-11 at 14.22.03.png and no idea which site any of them came
from, what fonts they used, or how the layout worked.
Nisaba is the browser that fixes that. You browse normally — but every capture keeps its source URL, its DOM, its CSS, its fonts, its design tokens and its provenance. The library gets more useful the longer you use it, and when you want the pattern back, an agent turns it into real code in your stack.
It is not your daily browser. It's the one you open when your library needs to grow.
Browse — multi-tab browsing where every site runs in its own isolated view: no Node, no preload bridge, every permission denied by default.
Capture — the visible viewport, the whole scrollable page (via CDP, beyond the viewport), a dragged region, or a picked element. Annotate with rectangles, arrows, highlight, blur, text and numbered callouts, stored as editable vectors so the original PNG is never touched.
Extract — hover the page and the DOM outlines itself; arrow keys walk to parent, child or sibling. Click and Nisaba keeps a robust selector, sanitized HTML, computed styles, CSS custom properties, fonts, palette, assets and accessibility metadata — plus framework detection with a confidence score and the evidence behind it.
Audit — review a page, live or on localhost, by clicking your way down it and pinning a note to each thing that needs fixing. Every pin remembers the element, its computed styles and where it sits on the page, then greps your workspace to find the file that renders it. Export the lot as a folder of tasks an agent can work straight through.
Profile — measure the colours, type scale, spacing, radii, shadows, breakpoints and :root
variables a page actually uses, into an editable design.md and tokens.json. Observed values and
inferred ones stay labelled apart.
Refine with agent — a measured profile is a guess: the heuristics pick one button, one input and one card out of everything on the page, and sometimes that is a chat bubble rather than the CTA. Hand the profile to the Claude Code or Codex CLI you already have installed and it reads the screenshot and every sample Nisaba took, then corrects the spec. Its answer is validated and folded onto the measurement — it may fix what was interpreted, never what was observed.
Element Style Matrix — find the buttons, inputs, cards and badges on a page, collapse visually identical instances into variants, and screenshot each interaction state the page really declares rules for.
Organize — Captures, Sections, Elements, Design Systems, Audits, Resources, Sites and Bookmarks, all searchable, filterable and on your own disk. Export the whole library as a plain folder and import it back with IDs and relationships intact.
Convert — hand a saved section to the Claude Code or Codex CLI you already have installed. Nisaba writes a source package, shows you the resolved prompt and the exact folder before anything runs, streams the log, and records what was produced with a trail back to its sources.
Similarity — every image is perceptually hashed as it is saved, so the library can find near-duplicates and answer "what else looks like this" without leaving your machine.
Verify — a component can run your project's own lint, type, test and build scripts, in order, stopping at the first failure. It is only marked verified when they all pass, or when you explicitly override a failure.
Preview — start the workspace's dev server from inside Nisaba and open what it serves in a tab, next to the source capture it was built from.
SQLite — the library is a real database with full-text search, using Node's built-in driver.
No native module, nothing to rebuild per platform. An older JSON index is migrated on first run and
kept alongside as index.json.migrated.
macOS builds are published on the Releases page, and the app updates itself from there.
Windows and Linux aren't shipped yet, but nothing in Nisaba is macOS-specific — build it yourself with the steps in Develop below. There is no account, no sign-in and no server; the whole thing runs on your machine.
git clone https://github.com/WynterJones/Nisaba.git
cd Nisaba
npm install
npm run dev| Command | What it does |
|---|---|
npm run dev |
Dev build with hot reload |
npm run build |
Type-check and build to out/ |
npm run dist |
Build a local installer into dist/ |
npm run release |
Build all platforms and publish to GitHub Releases |
src/
main/
browser.ts Tab lifecycle and isolated WebContentsViews
capture.ts Viewport, full-page (CDP) and region screenshots
extract.ts The in-page selection overlay and artifact collector
design.ts Whole-page token measurement and design.md generation
audit.ts The page-review overlay and pin context collector
sourcemap.ts Grepping a workspace to find what renders an element
elements.ts Primitive detection and per-state capture
workspaces.ts Folder selection, probing and the write boundary
jobs.ts Prompt resolution, agent spawning, output detection
exporter.ts Portable library export and import
library.ts On-disk index, the nisaba:// asset protocol
agents.ts CLI discovery and version probing
preload/ The narrow typed bridge; the only surface the app UI can call
renderer/ React + Tailwind + shadcn/ui application UI
_PLANS/ Product requirements and design references
Three trust levels, and remote pages sit at the bottom of all of them:
- Main process — owns windows, the database, the filesystem, screenshots and any CLI it launches. Nothing else touches those.
- App renderer — the Nisaba UI. Talks to main through one typed IPC surface, never Node.
- Remote page views — every site you browse, in its own
WebContentsViewwithnodeIntegration: false,contextIsolation: true, sandbox on, no preload bridge and every permission denied by default.
Text scraped from a webpage is treated as data, never instructions — including when it is packaged up and handed to an agent. Every job runs with its workspace folder as the working directory, and the resolved prompt opens by telling the agent that everything in the source package is untrusted third-party content.
No account. No cloud. Your captures, prompts and generated code stay on your machine, and the only thing that ever leaves it is what you explicitly send through an agent CLI you installed and authenticated yourself. Diagnostics are opt-in and off.
Issues and pull requests are welcome. Phase-scoped work is easiest to review — check the current phase in Status and in the PRD before starting something large.
Nisaba was the Sumerian goddess of writing, accounting and the scribal arts — the one who kept the records. Fitting for a tool whose whole job is making sure what you found doesn't get lost.
MIT © Wynter Jones
Off they go with the evidence.