Introducing bolt.gives: Open-Source Agentic Coding With Any LLM

An open-source project called bolt.gives has launched under the MIT licence, offering an agentic coding platform that turns plain-English prompts into running web applications. The agent creates and modifies source files, installs dependencies, runs build commands, starts the dev server and verifies the app loads in Preview, with configurable model providers so self-hosters can connect their preferred LLM rather than being locked to one ecosystem. The project can be self-hosted on Ubuntu and, for a private single-owner install, runs without PostgreSQL or SMTP.

AI has changed the way we write software remarkably quickly. But I believe there is an important difference between an AI that can write some code and an AI system that can actually build something. That distinction is what led us to bolt.gives. bolt.gives is an open-source Agentic Coding platform that lets you describe an application in plain English and turns that request into a real, working web application. And we've released it under the MIT licence. From a Prompt to a Working Application You might type: Build a responsive appointment booking application for a mobile dog groomer. Include services, available time slots, customer details and a polished confirmation screen. bolt.gives doesn't simply respond with a large block of code. It works on the project. The agent creates and modifies the actual source files, installs dependencies, executes commands, starts the development server and checks whether the application is genuinely accessible before declaring the job complete. You then see the application running in Preview. If something isn't right, you simply continue the conversation: Make the booking screen more modern and add a dark mode. The agent works against the same project, files, runtime, history and Preview. That distinction is important to us. Generated code isn't the finished product. A working application is. Bring Your Own AI One of the principles behind bolt.gives is that an open-source coding platform shouldn't dictate which AI provider you have to use. The architecture supports configurable model providers, allowing self-hosters to connect their preferred supported provider rather than tying the development environment to one particular AI ecosystem. As AI models continue to improve, we think this separation becomes increasingly important. The coding environment should survive the model cycle. Agentic Rather Than Just Generative There is a lot happening between entering a prompt and seeing your application. bolt.gives can: create and modify real project files; install project dependencies; execute development and build commands; run the application; monitor Preview health; recover from common dependency, manifest, routing and startup failures; retain project context for subsequent prompts; work with connected Supabase projects; and publish supported projects to a shareable Cloudflare-backed address. The objective is to progressively remove the gap between describing software and running software. Reliability Matters One lesson we've learned while building bolt.gives is that generating code is often the easy part. Making the result reliably run is considerably harder. An AI model can produce perfectly plausible-looking source code while leaving you with a missing dependency, malformed manifest, broken import, incorrect route or development server that never actually starts. So we've adopted a fairly simple rule: Preview is the contract. bolt.gives doesn't regard the existence of generated files as proof that the task succeeded. The runtime checks that Preview can actually load. We've also been building automatic recovery around common failures so that the agent can attempt to repair the project rather than immediately handing the problem back to the user. Built for Self-Hosting I also wanted bolt.gives to be something developers could actually own. The project can be self-hosted on Ubuntu, and the repository contains the application itself, installer, runtime components, documentation and development environment. For a fresh installation, PostgreSQL isn't even mandatory. A private single-owner installation can be created without a database or SMTP server, while those components remain available for deployments that need them. Six Modules, One Platform As bolt.gives has grown, we've separated the codebase into six major areas: Core handles shared contracts, paths, security, logging and low-level utilities. Agent handles models, prompts, context, streams, tools, browsing and recovery. Runtime handles hosted commands, workspace synchronisation, Preview lifecycle, health and isolation. Project handles files, history, the Workbench, editor, terminal, actions and integrations. Control Plane handles profiles, instances, publishing, billing, mail and auditing. Surfaces handles routes, application chrome, Cloudflare adapters and orchestration. The separation matters because we want contributors to be able to work on one part of the system without first understanding the entire platform. And It's Open Source This is perhaps the part I'm most excited about. bolt.gives is available on GitHub under the MIT licence. We aren't publishing a crippled demo repository whose interesting parts live somewhere else. We want developers to explore it, break it, improve it, fork it and help us decide where it goes next. There are already hundreds of commits behind the project, but we're still early in what we want bolt.gives to become. And that is exactly why I'm introducing it here on DEV. We Want Contributors There are many directions in which an agentic development environment can evolve. Better agents. Better context management. More frameworks. Better recovery. New model providers. Improved runtimes. Better deployment. Developer tooling. UI and UX improvements. Performance optimisation. And ideas we haven't thought of yet. If any of those areas sound interesting, we'd genuinely love to have you involved. Contributions use the familiar GitHub workflow: fork the repository, make your change and submit a pull request. Try It You can explore the source code here: GitHub: https://github.com/embire2/bolt.gives https://github.com/embire2/bolt.gives And you can try bolt.gives here: bolt.gives: https://bolt.gives https://bolt.gives If you're interested in AI-assisted development, agent architecture, developer tooling or open source, I'd particularly love to hear what you think. Tell us what works. Tell us what doesn't. Open an issue. Send a pull request. Or fork the whole thing and build something we never anticipated. That's the point of open source.