# OpenRoom (VibeApps): A Browser Desktop Where an AI Agent Controls Every App

> OpenRoom, published by MiniMax AI under the working name VibeApps in the README, is a browser-based desktop environment where an AI agent operates built-in apps through natural language, all running locally with no backend and data stored in IndexedDB.

**MiniMax-AI/OpenRoom** — A browser-based desktop where AI Agent operates every app through natural language.

- Repository: https://github.com/MiniMax-AI/OpenRoom
- Website: https://www.openroom.ai
- Stars: 1,267 · Forks: 166
- Language: TypeScript
- License: MIT
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/minimax-ai-openroom

## A browser desktop controlled entirely through natural language

OpenRoom presents a browser window that behaves like a desktop: resizable, draggable windows and a set of apps you can open side by side. The defining feature is an embedded AI agent that accepts natural-language instructions and operates the apps on your behalf. You do not click through menus; you type what you want and the agent figures out which app to address and what action to take.

The GitHub repository is MiniMax-AI/OpenRoom, and the homepage is https://www.openroom.ai. The README uses the project name VibeApps, which appears to be the application's display name. The target audience is developers who want to explore LLM-driven UI interaction, or who want a scaffold for building new browser apps with an agent-accessible Action interface. All computation runs in the browser. No account is required, and no data leaves the local machine; user data is stored in IndexedDB.

## Nine built-in apps and the Action system

Out of the box, OpenRoom ships nine apps: Music (full player with playlists and album art), Chess (with complete rule enforcement), Gomoku, FreeCell, Email (inbox, sent, drafts), Diary (journal with mood tracking), Twitter (a local social feed), Album (photo browser), and CyberNews (a news aggregator). Each app is fully integrated with the AI agent through a structured Action system.

Each app exposes a set of typed actions in an actions/ directory, with an APP_ID and action type definitions in constants.ts. The agent sends actions to a specific app using the app's APP_ID, and the app's state management layer (built with React Context and Reducer) handles the action. This explicit structure means new apps can be added without modifying the agent itself; the agent discovers available actions through the app registry. The in-browser chat panel is for controlling existing apps. Creating new apps requires Claude Code and happens in a terminal, not in the browser.

## Getting OpenRoom running locally

The README lists Node.js 18 or later and pnpm 9 or later as prerequisites. After cloning and installing dependencies, an optional environment file must be copied before the development server starts:

```bash
git clone https://github.com/MiniMax-AI/OpenRoom.git
cd OpenRoom
pnpm install
cp apps/webuiapps/.env.example apps/webuiapps/.env
pnpm dev
```

Open http://localhost:3000 to see the desktop. Double-click any app icon to open it. The chat panel appears by clicking the icon in the bottom-right corner. You will need to configure an LLM API key in the chat panel settings before the agent can respond. The README does not specify which LLM providers are supported, but the .env.example file is the place to look for required keys. For users in China, the README notes that mirror lines in .npmrc can be uncommented for faster downloads via npmmirror.

## The Vibe Workflow for generating new apps

OpenRoom ships a workflow engine at .claude/commands/vibe.md that Claude Code can run to generate a complete new app. You describe what you want in the Claude Code CLI, and the workflow runs through six stages:

```bash
/vibe WeatherApp Create a weather dashboard with 5-day forecasts and temperature charts
```

The six stages are: Requirement Analysis, Architecture Design, Task Planning, Code Generation, Asset Generation, and Project Integration. At the end, the new app is registered in the desktop and accessible from the app list. The workflow also supports evolving an existing app:

```bash
/vibe MusicApp Add a lyrics panel that shows synced lyrics during playback
```

That triggers a focused four-stage change workflow: Impact Analysis, Planning, Implementation, and Verification. Both workflows support resuming from a specific stage:

```bash
/vibe MyApp --from=04-codegen
```

This is useful when code generation succeeded but asset generation failed, or when a stage needs to be rerun without repeating earlier analysis steps.

## Project layout and the open-source vibe-container mock

The repository is a pnpm monorepo managed with Turborepo. The main desktop application lives in apps/webuiapps/. Each app's source sits in a subdirectory under apps/webuiapps/src/pages/, following a consistent structure: components/, data/ (seed data as JSON), store/ (Context + Reducer), actions/, i18n/ (English and Chinese translations), and meta/ (metadata for the Vibe workflow).

The packages/vibe-container/ directory is important: in the full commercial version of the platform, this package provides an iframe communication SDK. In the open-source standalone version, it is replaced by a local mock in src/lib/vibeContainerMock.ts that uses IndexedDB for storage and a local event bus for agent communication. The README explicitly notes that the type definitions and client-side SDK interface are still present, but the actual iframe communication layer is stubbed out. This means the open-source version is a self-contained prototype, not the production infrastructure used on openroom.ai.

## Limitations and what the agent cannot do

The agent can only operate apps that expose an Action system. Apps outside the defined action set cannot be controlled through natural language. There is no way to ask the agent to interact with an arbitrary web page or browser tab; the desktop is a closed environment containing only the registered apps.

The Vibe Workflow for generating new apps requires Claude Code (the Anthropic CLI) to be installed and authenticated. The workflow runs in the terminal, not in the browser. Teams using a different AI coding environment would need to adapt the .claude/commands/vibe.md workflow definition. The last push to the repository was on 2026-06-03, and the project has no official GitHub releases. The version in package.json is 0.1.0, marking it as early-stage. Teams should treat the current state as a developer preview rather than a stable platform.

## Comparison with Electron-based desktop applications

Electron wraps a web application in a native process with access to the local filesystem, native menus, and OS notifications. OpenRoom runs entirely in a browser tab with no native process. The tradeoff is that OpenRoom is more portable (any browser, any OS, no installation), but has narrower system access: data is limited to IndexedDB, there is no native filesystem access, and there are no OS-level integrations.

Frameworks like Tauri pursue a similar browser-plus-native approach but with a compiled Rust backend for system access. OpenRoom takes the opposite direction: pure browser, no backend, local-only data. This makes it straightforward to deploy as a static site, but it means the agent cannot read or write files outside the browser sandbox. A team building productivity tools that need file access or system integration would need Electron or Tauri; OpenRoom is suited for self-contained browser experiences where localStorage and IndexedDB are sufficient.

## Maintenance status and MIT license

The last push to MiniMax-AI/OpenRoom was on 2026-06-03. The project is not archived and carries an MIT license. No GitHub releases have been published; the version in package.json is 0.1.0. The MIT terms allow copying and embedding the code in other projects without restriction.

The repository includes CODE_OF_CONDUCT.md, CONTRIBUTING.md, SECURITY.md, and a CHANGELOG.md. Playwright is configured for end-to-end testing. The presence of a .takt/ directory at the root (though this is the OpenRoom repository, not the takt tool) suggests some internal workflow tooling. NOTICE and AGENTS.md are also present. The .claude/ directory at the root contains the Vibe Workflow commands. Given the 0.1.0 version and the approximately four-month gap since the last push, teams adopting this project should expect active development to resume or should fork and maintain their own version.

## Conclusion

OpenRoom is a reasonable starting point for developers who want to experiment with an AI agent that operates a set of pre-built browser apps, or who want a scaffold for building new apps that expose an Action API to an agent. It is not suited for teams that need a production-ready app platform; the last push was on 2026-06-03 and the repository has no official releases. The open-source vibe-container is a local mock of the commercial SDK, not the full platform. Before using the Vibe Workflow to generate new apps, confirm that Claude Code is installed and that a valid LLM API key is configured in the chat panel settings.

## FAQ

### What is OpenRoom AI?

OpenRoom AI (published as VibeApps in the README) is a browser-based desktop environment built by MiniMax AI where an embedded AI agent operates a suite of built-in apps through natural-language instructions, all running locally in the browser with data stored in IndexedDB.

### Does OpenRoom require an LLM API key?

The README states that you need an LLM API key to use the AI agent in the chat panel. The key is configured in the chat panel settings after the development server is running. The specific providers supported are listed in the .env.example file.

### Is OpenRoom's vibe-container the full commercial SDK?

The README explicitly states that the open-source standalone version replaces the real iframe SDK with a local mock in src/lib/vibeContainerMock.ts. The mock uses IndexedDB and a local event bus rather than the full iframe communication layer used in the commercial platform.

## Sources

- [Issues](https://github.com/MiniMax-AI/OpenRoom/issues)
- [License: MIT](https://github.com/MiniMax-AI/OpenRoom/blob/main/LICENSE)
- [MiniMax-AI/OpenRoom on GitHub](https://github.com/MiniMax-AI/OpenRoom)
- [Project website](https://www.openroom.ai)
- [README](https://github.com/MiniMax-AI/OpenRoom/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/minimax-ai-openroom