· via Hacker News – Front Page (hnrss.org)
Onlook pairs live app previews with bring-your-own AI agents in open-source design tool
YC W25 startup Onlook has launched an open-source, local-first desktop app that renders your real app on a canvas and lets a connected Claude Code or ChatGPT/Codex agent edit the code.
A design canvas wired to real code
Onlook, a Y Combinator Winter 2025 startup, has launched on Hacker News with an open-source desktop app that displays a running web application on a visual canvas and lets an AI coding agent reshape it by editing the project's actual files. According to the project's GitHub README, posted alongside the launch, the tool is pitched as an open-source alternative to Figma, V0, Replit Agent, Base 44 and Webflow.
The premise is to close the gap between design work and shipped software. Rather than drawing static screens that someone later has to rebuild, Onlook opens any Git repository on the user's machine, starts that project's own dev server, and shows the live result in phone, tablet and desktop frames arranged on a free-form canvas. Every change lands in the repository itself, with Git keeping track of what moved.
Bring your own agent
Onlook ships no AI model of its own. Users connect a Claude Code or ChatGPT/Codex account they already have, and that agent does the work inside the project. Extra tools can be attached through MCP connectors, and several design tasks can run in parallel across repositories and pages.
The README describes a workflow in which you describe what you want in chat, answer the agent's clarifying questions, then click frames or individual elements to steer follow-up edits. The agent can also arrange the canvas for you, laying out variations, states and flows, and can work from reference images. When a design is ready, a single click opens a pull request, running checks through to merge, and entirely new codebases can be started from scratch inside the app.
Local-first by design
The app is desktop-based and local-first: code stays on the machine, and a build from source needs no Onlook account, cloud keys or environment file. Because Onlook simply runs whatever dev server the project already uses, it says it works with any web framework or language that server supports. Design work happens in a worktree copy, so the original folder is never touched.
For sketches outside a repository, the tool offers artifacts, quick HTML pieces that export as self-contained HTML, PNG, JPEG or PDF for handoff, review or sharing in tools like Slack or email.
Technically, the desktop app is built with Electron, React 19, React Flow, Zustand and CodeMirror, with Bun and Turborepo in the tooling; image exports use Electron's capturePage and PDFs use printToPDF.
Availability and fine print
Downloads currently cover macOS only, for both Apple Silicon and Intel. Windows is listed as coming soon behind a waitlist, and there is no Linux artifact while the team gauges demand. The code is Apache 2.0 licensed, but the README warns that the license covers source only: Onlook's icons come from Central Icons by Iconists under terms that do not extend to forks, and brand marks, installer art and sounds would need replacing in a fork.
Privacy-conscious users should also note that product usage analytics are on by default in the official download, though a source build asks first; telemetry can be switched off in settings or by setting STUDIO_TELEMETRY=0.
Why it matters
Design tooling is converging with AI coding agents, and commercial products such as V0 tend to bundle their own models behind subscriptions. Onlook bets on the opposite arrangement: keep the editor open source, let designers plug in the agent they already pay for, and keep every byte of code on the user's own disk. For teams wary of pasting a codebase into a vendor's cloud, that local-first, model-agnostic stance is a real differentiator, and the repo-based approach removes the classic handoff problem where a mockup has to be rebuilt as real UI.
The constraints are just as clear. A macOS-only binary limits reach, the asset licensing complicates true forkability, and the tool's usefulness still depends on the quality of whichever agent a user connects. As a signal of where design tools are heading, open editors orchestrating third-party models, the launch is worth watching either way.
- #open-source
- #design-tools
- #ai-agents
- #developer-tools
- #electron