Onlook Edits the Real Project, Not a Copy of the Design

Onlook is an open-source developer tool for designers, released under the Apache 2.0 license. Its GitHub repository presents it as an open-source alternative to Figma, V0, Replit Agent, Base 44 and Webflow.
The architecture is the distinctive part. According to the repository, Onlook opens any Git repository on the user's computer and runs the user's own dev server locally. Git is the version control system that records every change to a project's files, and the dev server is the local program that renders the site while it is being built. Rather than exporting a design file into a separate codebase, the tool works on the project directly. Users can connect their Claude Code or ChatGPT/Codex account to edit real files in that repo, so AI-driven changes land as ordinary diffs in the working tree.
Onlook's website describes the product as designing with AI on an infinite canvas. The site says users can craft, preview and iterate with AI to ship websites and prototypes, and that the native design tool features work 1:1 with code.
Output formats are also documented in the repository. Onlook can export artifacts as self-contained HTML, PNG, JPEG or PDF.
Platform support is narrower. The desktop app is available for macOS on Apple Silicon and Intel x64. The Windows version is listed as "coming soon", and the repository publishes no Linux artifact. Because the core loop runs against a local Git checkout and a local dev server, that gap matters to developers working outside macOS. Anyone building from source will work with the repository itself rather than a packaged binary.
The broader context here is a shift in where design work lives. Figma-style tools keep the source of truth in a proprietary canvas, and handing off to engineering is a translation step. Onlook's pitch removes that step by making the canvas a view onto the code. A designer's adjustment and an engineer's commit become the same kind of artifact, reviewable in the same pull request. Whether teams adopt that model will depend on how well visual edits round-trip through component structure, styling conventions and framework idioms in real projects.
The model access matters too. By its own description, Onlook does not supply the language model. It connects to the user's Claude Code or ChatGPT/Codex account, so the billing, rate limits and data-handling terms of those services become part of the workflow. Teams with procurement or data-residency constraints will need to settle that configuration before rollout.
Apache 2.0 is a permissive license. It allows use, modification and redistribution, including in commercial products, subject to its notice and patent terms. For a project competing with hosted design platforms, that lowers the cost of forking, self-hosting and embedding the tool in internal pipelines. The project's trajectory will also be shaped by whoever contributes to the repository, not only by the company behind it.
In my view, the most useful test for Onlook is ordinary use on a mid-sized front-end codebase, with a designer and a developer editing the same component tree on the same afternoon. If the local-first, Git-native model holds up there, it offers a concrete alternative to the design-to-code handoff. If it does not, the Windows and Linux gaps and the dependency on third-party AI accounts will be the first problems a team runs into.


