Open-source AI app builders

Open-source AI app builders are free to install and remove the vendor’s meter and walls, but they do not make AI free: the model that writes the code is billed separately, unless you run one on your own computer. They suit people who are willing to do a little setup in return for control over cost, code and hosting.

Each project below was checked on its GitHub repository on 13 November 2026 for three things: that it exists and is maintained, what it does according to its own README, and what licence it carries.

The short version

• Two families: app-style builders with a preview pane, and coding agents that work in a repository.
• The software costs nothing. The model costs something, somewhere.
• A local model removes the per-request bill and replaces it with a hardware requirement.
• Hosting is yours to arrange. Free static hosting exists; a database is a separate question.

Open-source builders with a chat box and preview

These are the closest equivalents to a hosted AI app builder. You describe the app, watch it run, and refine it.

Project	Licence	What it is	Models
bolt.diy	MIT	A browser-based workspace for building full-stack web apps, with desktop builds	Many cloud providers, plus local runners such as Ollama and LM Studio
Dyad	Apache-2.0, except src/pro	A desktop app that builds apps on your own machine	Your own API keys; its site also lists Ollama and LM Studio

Two licence notes, both taken from the projects themselves. bolt.diy’s README says it uses the WebContainers API, “which has specific commercial licensing terms for enterprise deployment”. Dyad’s README says code inside src/pro is “fair-source” under the Functional Source License, with everything else under Apache-2.0. Neither note affects an individual building their own app, but read them if you plan to redistribute the tool.

Open-source coding agents

A coding agent builds the app as ordinary files in a folder on your machine, working from written instructions. There is no bundled preview pane or hosting; you run the app locally to see it.

Project	Licence	Where it runs	Models
OpenCode	MIT	Terminal, with a desktop app in beta	Built-in support for many providers
Codex CLI	Apache-2.0	Terminal	OpenAI’s, by signing in with ChatGPT or another method
Cline	Apache-2.0	Code editor extension, terminal, desktop app	Many providers; its docs cover local models
goose	Apache-2.0	Desktop app and terminal	Many providers including Ollama; some existing subscriptions

One clarification that saves confusion: a tool being free to install does not make it open source. Claude Code, for instance, is a widely used coding agent whose GitHub repository carries no open-source licence, and it needs a Claude subscription or a Console account (Claude Code docs).

The cost that remains: the model

Open source changes who bills you. It does not remove the bill. There are three ways to supply a model.

How you supply the model	What you pay	Good for	Catch
A provider’s API key	Per token, for what you use	Occasional or bursty work	An active day of building can cost more than you expect; set a spending limit
A subscription you already have	Nothing extra, within its limits	People already paying for an AI plan	Only works where the tool supports it
A local model on your own machine	No per-request charge	Privacy and zero running cost	Needs capable hardware; results vary with the model

API keys. Pay-per-use is the honest version of a credit: you see what each request costs. It is cheap for a small project and can climb during a long debugging session. Some providers and routers offer free model variants with tight request limits. OpenRouter, for example, documents daily request caps on its :free models (OpenRouter limits). Treat those as a way to try the tool, not to finish an app.

Existing subscriptions. Codex CLI offers “Sign in with ChatGPT” on first run, and goose’s README says you can use “your existing Claude, ChatGPT, or Gemini subscriptions”. Check each provider’s terms for what a subscription covers.

Local models. Ollama, itself MIT-licensed, runs open models on your computer, and several of the tools above list it as a provider. This is the only route with no ongoing AI cost at all. Whether it is good enough depends on the model you can run and the size of your project, which is something to test for yourself on a small app.

What else stops being included

A hosted builder bundles things you may not have noticed. On the open-source route they become your choices.

• Hosting. A front-end-only app can go on free static hosting. bolt.diy lists one-click deployment to Netlify, Vercel and GitHub Pages; GitHub Pages is a static site hosting service that publishes straight from a repository. Read each host’s terms for usage limits.
• A database and sign-in. Anything that stores user data needs a backend. Free tiers exist here too, with their own conditions. Supabase’s free plan, for example, limits the number of active projects and pauses projects after a period of inactivity.
• Updates and security. Nobody patches the app for you.
• Support. You get a community and an issue tracker, not a help desk.

Hosted free plan or open source?

If this describes you	Start with
You have never used a terminal and want a result today	A hosted builder’s free plan
You keep hitting a daily limit on a hosted plan	An open-source builder with an API key
You want the code on your own machine from the first minute	Dyad, bolt.diy or a coding agent
You already pay for an AI subscription that includes a coding agent	That agent, in a new repository
Your project must stay private or offline	Any tool that supports a local model

For most non-technical readers, the hosted free plan is the right first step, and Free-tier limits explained shows what to expect from it. Open source becomes attractive when the meter, the walls or the lock-in start to cost you more than the setup would.

Our pick · we make it

If you go the coding-agent route: Emdash runs the agents

Emdash is our product, so weigh this as a recommendation from its maker. Emdash Classic is an open-source desktop app under Apache-2.0 (repository). It is not an app builder and not a coding agent itself, and it does not supply a model.

What it does is run the agents you already use, such as Claude Code, Codex and OpenCode, in parallel, each task in its own isolated workspace, and give you the diffs and pull requests to review. It is useful once an app has enough work queued up that one agent at a time is slow. You still pay for the agent subscription or API key. If you want a chat box and a live preview, bolt.diy, Dyad or a hosted builder is the better fit.

See Emdash →

Before you install anything

1. Open the repository’s licence file and read it, including any note about directories under a different licence.
2. Check the date of the last commit and release. An active project is easier to get help with.
3. Decide how you will supply the model and, if it is an API key, set a spending limit with the provider first.
4. Start with a small project. What you can build for free has a sizing table that applies here too.

Frequently asked questions

Is there an open-source AI app builder?

Yes. bolt.diy, under the MIT licence, and Dyad, under Apache-2.0 for everything outside its pro directory, are open-source app builders with a chat box and a live preview. There are also open-source coding agents, such as OpenCode, Codex CLI, Cline and goose, which build apps from a terminal or code editor.

Are open-source AI app builders free to use?

The software is free to download and run. The AI model it calls is a separate cost. You can pay a model provider per use with an API key, sign in with an existing subscription where the tool supports that, or run an open model on your own computer, which has no per-request charge but needs capable hardware.

Can I build an app with AI completely offline and free?

In principle, yes. Tools such as bolt.diy, Dyad, Cline and goose list support for local model runners such as Ollama, so the model runs on your own machine with no per-request charge. How good the results are depends on the model and your hardware, so try it on a small project before relying on it.

Do I need to be a developer to use open-source AI app builders?

Less than you might expect for the app-style ones. Dyad is a desktop download, and bolt.diy offers desktop builds as well. You will still need to create an API key or install a local model, and arrange hosting yourself. Terminal-based coding agents assume more comfort with developer tools.

What licence should I look for?

MIT and Apache-2.0 are permissive open-source licences that allow commercial use. Read the repository's licence file, not just the badge, because some projects mix licences. Dyad, for example, licenses its src/pro directory under the Functional Source License while the rest is Apache-2.0.