One screenshot in, every model at once, a wall of redesigns out.
Paste a screenshot, get a wall of directions back, every one real, working HTML you can open.
- Claude
- GPT
- Gemini
- DeepSeek
- Qwen
Drop. Pick. Run. Compare.
-
A1 · Input
Drop in a screenshot
One image, or a folder of shots.
-
A2 · Configure
Pick models and presets
Or write your own prompt.
-
A3 · Execute
Hit run
Every model fires in parallel.
-
A4 · Review
Judge the wall
The original first, every result beside it.
Every major model. Plus whatever you add.
Adding another is a JSON entry, not a code change.
| No. | Model | Sees images | Source |
|---|---|---|---|
| B1 | Claude | Native | Default config |
| B2 | GPT | Native | Default config |
| B3 | Gemini | Native | Default config |
| B4 | DeepSeek | Via auto-description | Default config |
| B5 | Qwen | Native | Default config |
| B6 | Your own | Either way | JSON config, no code |
B7 · Prompt index
Faithful Refresh · Bold Reimagine · Clarity & Accessibility · Conversion · Minimalist · Material 3 · Approachable · Playful · or write your own.
One admin dashboard. Eight directions.
All eight from one run, exactly as returned. Filter to compare, click to enlarge.
Showing 08 / 08 · filter · all
A real instrument under the paint.
Every result is a live page you can click around in. The control panel runs the whole show from one screen.
-
D·1
Side-by-side viewer
Filter by model or preset, preview from phone to desktop width.
-
D·2
Reference images
A look you like lends its mood and colors, never its layout.
-
D·3
Self-healing keys
Stack keys per provider. Failing ones get benched, then quietly retried.
-
D·4
Sandboxed rendering
Model HTML runs in a locked-down iframe. It cannot reach your data.
-
D·5
Runs locally
Localhost only. Keys and screenshots never leave, except to the APIs you chose.
-
D·6
Reproducible and scriptable
Every run is written to a manifest. Full CLI twin, MCP server for agents.
One file. No install.
Grab the build for your platform, run it, open 127.0.0.1:5178, drop in your API keys. Done.
v1.6.7 · checksums · all releases
$ bun install && npm run build
$ cp .env.example .env
$ bun run src/index.ts serve
→ RēDesign serving at http://127.0.0.1:5178- Dry run · add --mock, costs nothing
Where RēDesign fits.
Three honest ways to get from one screenshot to a redesign, and what each one actually gives you.
| No. | Approach | Models per run | Output | What you pay for |
|---|---|---|---|---|
| F1 | RēDesign | 5 in parallel, plus any you add | Real HTML per model, side by side | Free tool · your own API usage |
| F2 | A single AI chat (ChatGPT, Claude, Gemini, etc.) | 1 per conversation | One chat reply at a time | Whatever that chat product charges |
| F3 | Manual redesign (Figma and similar) | None | Hand-built mockup or frontend | Design tool subscription for team features, free tier for solo use · your time |
One AI chat means a new tab per model. RēDesign dispatches one input to every model at once, lined up against the original.
Questions, answered.
Is RēDesign free?
Yes, RēDesign is free to download and MIT licensed, with the source on GitHub. Grab a prebuilt release for Windows, Linux, or macOS on Apple silicon, or build it from source with Bun 1.2 or newer. There's no install step for the packaged build, just run it and open the control panel.
Which AI models does RēDesign support?
RēDesign ships with Claude, GPT, Gemini, DeepSeek, Qwen, and Meta AI configured by default, and runs them all in parallel on a single screenshot. Claude, GPT, Gemini, and Qwen see the image directly; DeepSeek gets an auto-generated description of it instead. You can add any other model as a JSON config entry, with no code change required.
Can I add my own AI model?
Yes. Add any other model, including your own, through a JSON config entry, no code change needed. It works whether the model can see images directly or only handle a text description. Config lives outside the app itself, so pointing it at a newer model version or disabling one doesn't require touching the app's code.
Does RēDesign keep my screenshots and API keys private?
Yes. RēDesign runs locally on localhost only, and your screenshots and API keys never leave your machine except to the model APIs you chose to use. Each model's generated HTML renders inside a locked-down iframe that cannot reach your data. Your keys are stored in a local, owner-only file and are never written into a run's saved output.
What operating systems does RēDesign run on?
RēDesign runs on Windows, Linux, and macOS on Apple silicon. On Windows there's a standalone GUI executable with the web app embedded, plus an optional system-tray build. On any supported OS you can instead run it from source with Bun 1.2 or newer, using bun install, a build step, and bun run src/index.ts serve.
Can I try RēDesign without spending on AI API calls?
Yes. Add the --mock flag to a run and it costs nothing, so you can try the flow before spending on any model's API. When you do run for real, RēDesign shows a cost estimate before you hit Run and tracks a per-run cost meter afterward, so a big fan-out across several models never surprises you.
Does RēDesign work offline?
The control panel itself runs entirely on localhost, no cloud account needed. Generating a real redesign needs internet access, though, since that step calls out to whichever AI provider APIs you've configured. A --mock dry run skips those calls entirely, so you can exercise the full flow without depending on any model being reachable.
How is RēDesign different from just asking ChatGPT or Claude directly?
Chatting with one model directly gives you one model's take per conversation, and comparing another means opening a new tab and re-pasting the screenshot. RēDesign fires the same screenshot at ten models in parallel by default, then lines every result up beside the original as real, working HTML you can open, filter, and click through, not just a chat reply.