- Home
- AI Computer Use Agents
- OpenComputer
OpenComputer
What it is
OpenComputer is a free, open-source desktop agent that runs locally, interprets the screen, and operates applications with mouse and keyboard actions. It can click, drag, scroll, enter text, use shortcuts, open URLs, and wait for interface changes. A vision-language model can run locally through vLLM or connect to a hosted API; the endpoint must accept image inputs, function tools, and the model’s tool-call format. Ask first is the default permission mode, requiring approval for mouse, keyboard, and URL actions, while Automatic mode removes individual prompts. The agent currently controls only the primary display. Linux X11 is best supported; Wayland, Windows, and macOS are experimental. Build targets include Linux AppImage and deb, Windows NSIS, and macOS DMG. Screenshots and prompts can remain local, but API keys are not stored in the operating-system keychain. The project describes itself as early alpha.
For and against
For
- Runs locally with screen understanding
- Supports mouse, keyboard, and URL actions
- Ask first approval mode is default
- Available for Linux, Windows, and macOS
- MIT-licensed open-source project
Against
- Project is described as early alpha
- Only the primary display is supported
- Wayland, Windows, and macOS are experimental
- API keys lack operating-system keychain storage
The verdict
OpenComputer offers local, screen-driven desktop control with explicit approval prompts by default. Its early-alpha status and experimental platform support call for supervision, especially for sensitive actions.
Who it suits
It suits users who want local desktop automation with screen understanding and configurable human approval before actions.
On the record
- Action review
- Ask first is the default permission mode and requires approval for mouse, keyboard, and URL actions; Automatic mode runs actions without individual prompts.github.com · read 7 Oct 2026
- API compatibility
- The configured endpoint must support image inputs, function tools, and the model’s tool-call format.github.com · read 7 Oct 2026
- Approvals
- Ask first is the default permission mode and requires approval for mouse, keyboard, and URL actions.github.com · read 7 Oct 2026
- Credential storage
- API keys are stored in the per-user application-data directory with user-only file permissions, but are not yet backed by the operating-system keychain.github.com · read 7 Oct 2026
- Desktop controls
- It supports clicking, dragging, scrolling, text entry, keyboard shortcuts, opening URLs, and waiting for UI changes.github.com · read 7 Oct 2026
- Display limit
- The current agent operates only the primary display, and multi-monitor automation is on the roadmap.github.com · read 7 Oct 2026
- Installer formats
- Build targets are Linux AppImage and deb, Windows NSIS, and macOS DMG.github.com · read 7 Oct 2026
- Integrations
- The current model boundary uses an OpenAI-compatible Chat Completions API; dedicated Discord, search, and MCP integrations are listed as future work.github.com · read 7 Oct 2026
- License
- OpenComputer is licensed under the MIT License, while third-party dependencies and model weights retain their own licenses.github.com · read 7 Oct 2026
- Maker
- The linked GitHub profile identifies the maker as Andrew Krikorian and lists California.github.com · read 7 Oct 2026
- Model compatibility
- The model endpoint must support image inputs, function tools, and the model's tool-call format.github.com · read 7 Oct 2026
- Model options
- It can use a local vision-language model through vLLM or a compatible hosted API.github.com · read 7 Oct 2026
- Privacy and security
- The README says screenshots, prompts, and tool results can stay on the user's machine, and that credentials, screenshots, model calls, and automation remain in Electron's main process.github.com · read 7 Oct 2026
- Project maturity
- The project describes itself as early alpha and warns that desktop actions should be supervised, especially around sensitive or irreversible workflows.github.com · read 7 Oct 2026
- Purpose
- OpenComputer is an open-source, local-first desktop agent that understands the screen and operates applications through mouse and keyboard interactions.github.com · read 7 Oct 2026
- Security design
- The Electron renderer is sandboxed with Node.js integration disabled, while privileged model and desktop automation behavior runs in the main process.github.com · read 7 Oct 2026
- Security limitations
- The project says its controls are defense in depth rather than a complete sandbox, and Automatic mode lets the model act with the user's desktop privileges.github.com · read 7 Oct 2026
- Security reporting
- The security policy asks users to report vulnerabilities privately and says reports should receive an initial response within seven days, though response times may vary because it is a volunteer project.github.com · read 7 Oct 2026
