The problem

AI agents get UI ~90% of the way there — then stall. The layout looks right, but a card is 8px too tall, a panel is the wrong shade of gray, a shadow is flat, a gap is off by 24px. Asking the model to "look at the screenshot again and fix it" is slow, and you get a different answer every time.

one-shot-ui closes that last 10% deterministically. It extracts structured data from a reference screenshot — layout regions, colors, typography, spacing, design tokens — diffs your implementation against it, and returns specific, ranked fixes, not "make it look more like this."

Set width to 616px (currently 640px)
Change the fill color to #303040.
Set box-shadow to -3px 0px 24px 0px rgba(28, 29, 38, 0.32).
gap: 176px; /* currently ~152px */

Copy-paste CSS, ranked by visual impact — every example above is real output from the run below.

Watch it converge

The run command loops extract → capture → compare → fix until the heatmap goes quiet. In the run above, the agent's first build looked identical to the eye — one-shot-ui flagged 15 concrete deltas (position, size, color, shadow, spacing) and the loop drove the build to ~2.5% pixel mismatch, within ~0.5% of the tool's own estimated irreducible floor (≈2%, sub-pixel font rendering) for this design.

Why one-shot-ui

• Deterministic, not vibes. Stable pixel + structural diff scores — same input, same numbers — so you can gate CI on "is this pixel-close enough?"
• Exact fixes, not nudges. It returns concrete CSS (width: 616px, #303040, gap: 176px), grouped by component and ranked by visual impact.
• Structural, not just pixels. Detects missing/extra elements, position & size shifts, color, shadow, spacing, and typography — and labels which differences are irreducible (anti-aliasing, photographic content) so agents don't chase ghosts.
• Not gameable. converge scores structural fidelity — is the reference's content present, placed, and legible? — not just pixel mismatch, and it refuses to win pixels by hiding, overlapping, or recoloring text into its background. It even flags reference text your build renders but clips out of view (a fixed-height overflow:hidden box) — the failure a pixel diff rewards, because cropped text lowers the mismatch.
• Agent-native. Ships an AGENTS.md (auto-discovered by Claude Code, Cursor, Codex, …) plus a Claude Code skill, so your agent drives it without hand-holding.
• Local & private. Pixel diffing, OCR, and layout extraction all run on your machine. No images leave your box, no API keys.

Copy a UI's style, not just match it

Matching a screenshot pixel-for-pixel is one job. The other: building a different screen that feels like it belongs to the same design system. one-shot-ui extracts the design language from a reference and verifies a new UI conforms to it — deterministically, with no pixel oracle to lean on.

# Extract a reusable style system — palette, spacing scale, type ratio, radii, elevation
one-shot-ui tokens reference.png --emit shadcn      # or: tailwind | json

# Build a different screen in that style, then check it conforms
one-shot-ui style-check reference.png ./pricing.html

It reports measured facts and leaves the taste to the agent: it won't name your "primary" color or classify the mood — you decide those from the image (you're the vision model; the tool is your exact, hallucination-free eyes). Conformance is judged on what a screenshot grounds reliably — palette, spacing rhythm, type scale — while roundedness and elevation are flagged as advisories, since a raster can't pin them down.

Install

npm install -g one-shot-ui

For commands that need a browser (capture, run):

npx playwright install chromium

Quick start

# Diff your implementation against a reference and see exactly what's off
one-shot-ui compare reference.png build.png --json --heatmap heatmap.png

# Get copy-paste CSS fixes, ranked by impact
one-shot-ui suggest-fixes reference.png build.png --json

# The step that gets you pixel-perfect: trial fixes in a live browser and keep
# only the ones that provably reduce pixel mismatch — outputs a verified patch,
# a structural fidelity score, and any reference text your build hides or clips
one-shot-ui converge reference.png --impl ./index.html

# Or run the full automated loop until it converges
one-shot-ui run reference.png --impl ./index.html --max-passes 5 --threshold 0.02

Every command supports --json for structured, agent-friendly output.

Use it with your coding agent

one-shot-ui ships an AGENTS.md (auto-discovered by Claude Code, Cursor, Codex, and other agent tools) plus a skill/SKILL.md for Claude Code.

Install the skill in one line:

mkdir -p .claude/skills/one-shot-ui && cp "$(npm root -g)/one-shot-ui/skill/SKILL.md" .claude/skills/one-shot-ui/

Use it as an MCP server

one-shot-ui also runs as a local MCP server, so any MCP-capable agent (Claude Code, Cursor, Cline, Windsurf, VS Code) can call compare, converge, suggest_fixes, extract, tokens, plan, and style_check as tools — no shell glue. It runs over stdio, makes no network calls, and needs no API keys.

claude mcp add one-shot-ui -- npx -y one-shot-ui mcp

Or add to any client's MCP config:

{
  "mcpServers": {
    "one-shot-ui": {
      "command": "npx",
      "args": ["-y", "one-shot-ui", "mcp"]
    }
  }
}

See docs/MCP.md for per-client setup and registry publishing.

Commands

Command	Purpose	Key Flags
extract	Analyze a screenshot into layout, color, and text data	--json, --no-ocr, --overlay, --fine
compare	Pixel + structural diff between two screenshots	--json, --heatmap, --dom-diff
tokens	Design tokens + a reusable style system (palette, spacing, type, radii)	--json, --emit shadcn|tailwind|json
style-check	Check a new UI conforms to a reference's design language	--json (new UI = URL / HTML / screenshot)
plan	Generate an implementation strategy	--json
capture	Screenshot a URL or local HTML file	--url, --file, --output
suggest-fixes	Tailwind/CSS fix suggestions from a diff	--json, --top, --dom-diff, --framework
converge	Closed-loop optimizer: pixel-verified CSS patch + structural fidelity score, flags hidden/clipped text	--impl, --out, --json, --budget-seconds
run	Multi-pass extract→capture→compare→fix loop	--impl, --max-passes, --threshold
benchmark	Run benchmark suites	--json, --output

How it works

1. extract — segments the reference into layout regions, samples colors/tokens, and OCRs text.
2. capture — screenshots your implementation (URL or local HTML) at a matched viewport.
3. compare — aligns the two, computes a pixel heatmap and a structural diff, and classifies each issue (layout / color / typography / spacing) plus whether it's actionable.
4. suggest-fixes — turns issues into concrete, ranked CSS edits.

run chains all four in a loop until the diff drops below --threshold.

Development

Requires Bun.

bun install
bun run install:browsers   # Playwright Chromium
bun run typecheck

Dev scripts run directly from source:

bun run dev:extract -- ./reference.png --json
bun run dev:compare -- ./reference.png ./build.png --json

Build for npm:

bun run build

License

MIT