Open-source Lovable alternatives

Onlook is an open-source, visual-first code editor for building websites, prototypes and designs in Next.js and Tailwind CSS. Edits are made directly on the browser DOM with a Figma-like interface, while the changes are written to real code, so design and development share one source of truth.

You can create a Next.js app from text or an image or start from templates, then edit it visually, manage brand assets and tokens, create pages, browse layers and manage images. Developer tools include a real-time code editor, checkpoints you can restore, branching for design experiments, a CLI, an app marketplace, one-click deploys, shareable links and custom domains. Team collaboration is on the roadmap.

The README describes this repository as the open-source editor that started Onlook, while the hosted product for AI-native designers is in early access through a waitlist. It is licensed under Apache-2.0 and presented as an open alternative to tools such as Webflow, Figma Make, V0, Replit Agent, Lovable and Bolt.new. It suits designers who want to ship real code and developers who want a visual layer.

Magic is an open-source platform for generating full-stack applications and AI agents from natural language. It produces the database layer, a secured API, business logic and a frontend, and positions itself as an alternative to Lovable, Bolt and Replit that gives you the whole backend on your own hardware without lock-in.

An API wizard can point at an existing MySQL, PostgreSQL, SQL Server or MariaDB database and generate a complete secured REST API, which you can then extend in plain English without a build or deploy step. It can wrap third-party APIs from an OpenAPI or Swagger URL, build database-driven agents behind role-based access control, create embeddable chatbots grounded in crawled website content, run scheduled jobs and serve static sites and single-page apps. An MCP server exposes every endpoint as a tool for Claude, Cursor or Codex.

Magic is built on .NET with the Hyperlambda language and is licensed under MIT. It runs locally in about a minute using Docker, and its stated use cases include CRM systems, admin panels, booking systems and internal tools. It suits developers who want AI-generated apps with a self-hosted, auditable backend.

Sandboxd is an open-source, self-hosted take on prompt-to-app builders like Lovable, Bolt, v0 and Replit. A single HTTP request creates a private, isolated container, runs an AI coding agent in it against your prompt, and returns a live preview URL for the resulting app, all on a server you own.

Beyond building from a prompt, it can also run existing apps from a curated catalog of more than 80 apps such as n8n, Ghost and Grafana, and every action is exposed as a /v1 API call for headless, scriptable use. Idle sandboxes go to sleep and wake on demand, so one ordinary server can hold many apps rather than a virtual machine each.

The design is intentionally small: a single Go program drives Docker, uses Traefik to route preview URLs and keeps state in SQLite, with no Kubernetes, separate database or queue. Self-hosting is free under the MIT license, and the maintainers also mention an optional managed cloud and a live demo console with sample data.

Micracode is an open-source AI web app builder presented as an alternative to tools like Lovable, v0, Bolt and Replit. You describe an app in natural language and it streams code into an in-browser workspace, where you can keep iterating by chat or edit files directly in a Monaco editor and preview the result live. It is aimed at developers and makers who want AI app generation to run on their own computer.

It ships as a desktop app, with a macOS build on the Releases page, and its core backend runs in-process, so there is no separate server to start. On first launch you paste an API key for OpenAI, Google Gemini or Anthropic, and the keys are stored in a local config file; the README says nothing is sent to a Micracode server. Projects are saved as plain folders on disk. There is no database and no sign-in, and cloud sharing happens only on request.

An optional feature publishes a temporary public static preview link through temp.md, which can be updated or revoked from the workspace. The codebase uses TypeScript, Electron and the pi coding agent SDK, and the project is licensed under MIT. Model usage is billed by whichever provider's key you use.

Cofounder is an experimental AI framework for generating full-stack web applications. Given a description of what to build, it aims to produce not just a front end but a working backend, database schema and stateful web app, pairing an AI-guided mockup designer with a modular design system to create a generative UI rooted in the app's own architecture.

The stack it generates targets TypeScript and JavaScript with React, Tailwind CSS, Vite and PostgreSQL. The README is unusually blunt about the project's maturity: it is an early, unstable preview release that is expected to break often and consumes a large number of AI tokens, and the maintainers explicitly tell most potential users to wait for a version 1 release rather than try the current code.

Cofounder is released under the MIT license and is run locally or self-hosted, calling out to an AI provider such as OpenAI for generation, so API usage is billed separately by that provider. It is best suited to developers comfortable with experimental, work-in-progress open-source tools.