// the find
stackblitz/bolt.new
Prompt, run, edit, and deploy full-stack web applications. -- bolt.new -- Help Center: https://support.bolt.new/ -- Community Support: https://discord.com/invite/stackblitz
Bolt.new is the open-source version of StackBlitz's AI coding agent — it runs a full Node.js environment in the browser via WebContainers and lets an LLM scaffold, edit, and run apps without any local setup. It's aimed at developers who want to self-host or extend the bolt.new product, not just use it at the URL.
The WebContainers integration is the real differentiator — you get actual npm installs and a running Node server in-browser, not a sandboxed code editor. The action/artifact streaming architecture in message-parser.ts is well-thought-out: the LLM emits structured actions (file writes, shell commands) that the runtime executes incrementally. State management via nanostores keeps the client-side complexity manageable. The file diffing and workbench store separation means the editor panel doesn't need to reload the whole preview on every keystroke.
The repo is effectively a snapshot of an internal product at one point in time (last push Dec 2024) — the community fork ecosystem has already diverged so far that upstream changes here may not reflect what people are actually running. There's no backend persistence layer in this codebase; conversation history is IndexedDB only, which means any serious self-hosted deployment needs to bolt on its own storage. The system prompt in prompts.ts is doing a lot of heavy lifting with no way to configure it at runtime — swapping models or adjusting behavior requires code changes. Test coverage is thin: one message-parser spec and CI that just checks the build, so regressions in the action-parsing logic are easy to miss.