A native desktop app for AI agents that work on your computer - with any model provider and your own API keys.
Like Codex or Co-Work, but provider-agnostic: OpenAI, Anthropic, Google, local Ollama models or any OpenAI-compatible endpoint, all from one window. No silos, no vendor lock-in.
- Chat and code in your projects - the agent runs in the project directory you pick, with file tools sandboxed to it.
- Give agents their own computer - run an agent on a Kubernetes workspace, watch its screen inside its chat or underComputers , and take over the mouse and keyboard when it gets stuck. Each workspace has its ownComputer use ,Browser use ,Recording andText to speech switches in Settings > Workspaces, on by default - the ones in Settings > General, and the speech one in Settings > Voice samples, cover this computer only. A workspace can also mountvolumes - existing claims of its namespace, read-only where the agent should copy before it edits - so a heavy app or raw footage is not downloaded to every machine. On the local cluster, Settings > Compute sharesfolders of this computer with its workspaces - mounted straight from here with no copy, read-only for every workspace or for one.
- Drive your browser - theopentask extension lets the agent navigate, click, type and read tabs in your everyday browser.
- Write your own agents - Markdown files with a name, instructions, model and face. Agents can hand work to each other.
- Make content - talking-avatar clips, text to speech and text to video on a timeline in Content projects.
- Stay portable - the app updates itself, and Export / Import moves your whole setup to a new machine.
Download the asset for your platform from the latest release:
| Platform | Download |
|---|---|
| macOS (Apple Silicon and Intel) | .dmg |
| Linux x86_64 | .AppImage ,.deb or.rpm |
| Windows x64 | -setup.exe or.msi |
Releases are not signed by Apple or Microsoft, so the first launch needs one extra click: on macOS right-click > Open and confirm, on Windows choose "More info" > "Run anyway". On first run the app downloads the infer CLI and the gateway into ~/.infer/ (on Windows you install the gateway yourself, see How it works). Then add your provider keys under Settings > API Keys and start chatting.
- [How it works](https://github.com/inference-gateway/desktop/blob/main/docs/how-it-works.md) - the gateway and CLI, working directory and sandbox, updates, supported platforms.
- [Agents and computers](https://github.com/inference-gateway/desktop/blob/main/docs/agents.md) - write your own agents and give them a machine of their own.
- [The Mac's GPU](https://github.com/inference-gateway/desktop/blob/main/docs/gpu.md) - run the local cluster on Podman so its computers use this Mac's GPU.
- [Browser use](https://github.com/inference-gateway/desktop/blob/main/docs/browser-use.md) - connect the opentask extension.
- [Avatars](https://github.com/inference-gateway/desktop/blob/main/docs/avatars.md) - lip-synced talking clips of yourself in Content projects.
- [Export / Import](https://github.com/inference-gateway/desktop/blob/main/docs/export-import.md) - move your setup to a new machine.
- [examples/computer-use](https://github.com/inference-gateway/desktop/blob/main/examples/computer-use) - three agents on Ubuntu desktops in a local Kubernetes cluster.
- examples/filmcraft - an editor agent with FilmCraft on every computer, from a folder shared with the local cluster.
- examples/drone-footage - an agent edits raw drone footage end to end: stabilisation, a color grade, ElevenLabs sound and one export per platform shape.
Built with Tauri and a React + TypeScript frontend (Bun + Vite, Tailwind v4 + shadcn/ui), powered by Inference Gateway. See CONTRIBUTING.md for prerequisites, local setup and building.