# Ozone Describe an app and Ozone builds it, runs it and shows it to you. It creates the app from a tested template, starts it with Docker Compose, hands the work to coding agents, checks every step itself and opens a live preview beside the conversation. It also has a plain chat that takes files. Ozone is local first: it runs on your machine with your own provider accounts, and sends nothing to a service you did not set up. Under the app builder is a full client for coding agents, with threads, providers, checkpoints, a preview browser and remote access over your own network. Ozone is built on [T3 Code](https://github.com/pingdotgg/t3code) by T3 Tools. See [Thanks](#thanks). ![Home: describe an app, pick a template, and see the apps you have built](docs/ozone/assets/home.png) ![A build thread: the plan with its repair task, the running app and its preview](docs/ozone/assets/build.png) That thread started from one sentence: "A reading list. Add a book with its title and author, mark it as read, and see how many are left to read." This is the app it built: ![The reading list app that the build produced](docs/ozone/assets/app.png) ## What it does - **Builds apps from a prompt.** A new app is a real project on your disk, in a git repository of its own, started from a tested template: a **web app** (TanStack Start with React 19, Tailwind CSS 4, Drizzle on Postgres and Better Auth), a **landing page** with no database, or an **API** (Hono, Drizzle on Postgres and zod). Apps with pages carry design and SEO skills for the agents, and a landing page's final check covers every page's title, description and place in the sitemap. - **Runs them in Docker.** Each app is a Docker Compose project with one address that never changes. Start, stop, restart, rebuild and read its logs from the thread. Give it settings and secrets under **Environment**: they reach its container and never its code. - **Orchestrates the work.** The Build Conductor turns a request into a task graph. A worker does each task. The server, not the agent, runs the type check, the linter and the tests inside the app's container. A failed check goes to a fresh worker with the evidence. Every task that passes is a commit. When every check passes, a tester uses the app in a browser against what you asked for, at phone width and in dark and light too. - **Gives each app a page.** Open it from Home or from the app's thread: Overview, Data, History, Environment, Build and Logs while you build, then Deploy, Deployments and Monitoring for production. It is laid out like Settings. - **Lets you steer.** Turn on "Plan first", on Home or above the message box in an app's thread, and nothing is built until you approve the plan: the agent asks its questions and shows the plan on cards in the conversation. With Claude, Codex, Cursor and OpenCode it cannot change a file while it plans. Turn on "Pick the design" to choose from three design directions, drawn as previews, before anything is built. Each app's Build page chooses the model its workers use, so a large lead can direct cheaper workers. - **Keeps your data.** An app's database lives in a volume that only a confirmed **Delete app** removes, after a snapshot, and Home can bring a deleted app back with that data. Every finished task takes a snapshot, and **History** takes the code and the data back to any task, which you can undo the same way. A migration that can lose data waits until you apply it, after a snapshot. - **Deploys.** **Deploy** runs the production build with a database of its own, on this machine or on a server that runs Docker, reached over SSH. Your development data is never touched, and Stop keeps the production data for the next deploy. Each deploy saves the production data first, and **Data** backs it up and restores it. - Give it a **domain** and it is served over HTTPS with a certificate from Let's Encrypt, and each new version takes over without the app going down. - **Deployments** lists every deploy with its output, and rolls back to any of them, with the data from then if you ask. - **Monitoring** shows each container's CPU, memory and network, and its logs. **Set up** checks a server, installs Docker there if it has none, and starts the proxy. - A **Dokploy** server, or a **Railway**, **Render**, **Koyeb** or **Vercel** account, can run the app instead. Add it once under **Servers** with an API key or token. The host then owns the containers and the data. Vercel takes only apps without a database. - Those hosts build from a Git repository. When the app has none, **Deploy** offers to create a private or public GitHub repository for it in one click. - **Build** deploys by itself when a build completes if you ask, **Deploy** gives the app a URL to call from a script, and **Notifications** tells Slack, Discord, Telegram or any webhook how each deploy went. - **Chats.** Ask anything and attach PDFs, Word documents, images and other files. Ask a chat for an app and it starts one. - **Uses your subscription.** Sign in to Claude from Settings. No API key is involved, and an API key left in your shell is not used by accident. Codex, Cursor, OpenCode and the other [supported providers](docs/README.md) work too. ![The Deploy page: pick a saved server, and create the app's GitHub repository in one click](docs/ozone/assets/deploy.png) ## Run it You need Docker with Compose, and [Claude Code](https://claude.com/product/claude-code) or another supported provider on the same machine. Ozone tells you on its home screen when one is missing, and how to get it. ### Install the desktop app Download the latest build from [Releases](https://github.com/koneb71/ozone/releases/latest): | Your computer | File | | ---------------------------- | ------------------------- | | Mac with Apple silicon | `Ozone-…-arm64.dmg` | | Linux, 64-bit Intel or AMD | `Ozone-…-x86_64.AppImage` | | Windows, 64-bit Intel or AMD | `Ozone-…-x64.exe` | The Mac app is signed and notarized by Apple, so it opens like any other download. The Windows installer is not signed: SmartScreen warns before it runs; choose **More info**, then **Run anyway**. Ozone keeps its data in `~/.ozone`. See [Install Ozone](docs/user/install.md) for the `ozone` command and for running it on a server. If no release is published yet, build the installers yourself: on GitHub, open **Actions**, choose **Ozone release** and press **Run workflow**. The files are attached to the finished run. ### Run from source Ozone builds with [Vite+](https://viteplus.dev/guide/), which also provides the right Node version. ```sh curl -fsSL https://vite.plus | bash vp i vp run dev ``` Open the pairing URL the dev runner prints. ### Build your first app 1. Open **Settings**, choose **Providers** and sign in to Claude. 2. On the home screen, describe an app and press Enter. 3. Watch the plan run. The preview opens by itself when the app is up. ## How it works - [Apps and builds](docs/user/apps.md) and [Deploying an app](docs/user/deploy.md): the guides for what Ozone adds. - [Using Ozone](docs/README.md): the rest of the user guides, from the composer to remote access. - [How Ozone is built](docs/ozone/README.md): how the parts fit and the decisions behind them. - [The Build Conductor](docs/ozone/orchestrator.md): the orchestrator and the decisions behind it. - [Data safety](docs/ozone/data-safety.md): the contract that keeps an app's data. - [AGENTS.md](AGENTS.md) and [CONTRIBUTING.md](CONTRIBUTING.md): how to work on Ozone. ## Thanks Ozone started as a fork of [T3 Code](https://github.com/pingdotgg/t3code), the open source client for coding agents by Theo and the team at T3 Tools. The threads, providers, checkpoints, preview browser, terminal and most of what makes Ozone fast and pleasant to use are their work. Ozone adds the app builder, the deploys and a local-first setup on top, and has since gone its own way. If you want a great client for coding agents without the app builder, use T3 Code. ## License MIT, as T3 Code is. See [LICENSE](LICENSE), which keeps T3 Tools' copyright notice beside Ozone's.