What is OpenComputer?
OpenComputer, by the developer andykr1k, is an Electron app for Linux, Windows and macOS that sits on screen as a small orb. You type a request; it takes a screenshot, sends it to a vision-language model, and carries out the next step with real mouse movement, clicks, drags, scrolling and keystrokes in ordinary desktop applications, taking a fresh screenshot after every action.
It talks to any OpenAI-compatible endpoint that accepts images and tool calls. The default is RedHatAI/Qwen3.5-9B-quantized.w4a16 served locally through vLLM, chosen to fit many 16 GB GPUs, so screenshots need never leave the machine. The project is MIT licensed and was published as a single commit on 22 July 2026 with no tagged releases. It is not related to Digger's cloud agent of the same name at opencomputer.dev.
Is OpenComputer worth it?
OpenComputer does exactly one thing: it lets a model see your screen and operate your mouse and keyboard. That makes it a pure computer-use agent, with none of the file editors, connectors or multi-agent planning of Eigent or OpenWork, and its value is in how plainly it does that job locally. With the default 9B quantised Qwen model on a 16 GB GPU, the screenshots, prompts and actions stay on your machine.
The safety design is sensible for an alpha. The default Ask first mode takes screenshots automatically but requires approval for every mouse, keyboard and URL action, with Enter to approve and Delete to deny; Automatic mode runs without prompts. Coordinates and URL schemes are validated, runs stop after 25 steps by default, and a stop control cancels a run in progress. The renderer has no Node.js access, so credentials and automation stay in the main process. The README says outright that these controls reduce risk but do not make model-driven desktop automation safe, and that text on any web page or document can carry a prompt injection.
The project's state is the problem. The repository holds a single commit from 22 July 2026, has 31 stars, three open issues and no tagged releases, so there are no installers: you build from source with Node 22.12 or newer and run the model yourself. Only Linux under X11 is listed as best supported; Wayland, Windows and macOS are experimental, the agent sees the primary display only, and API keys sit in a user-only file rather than the system keychain.
For a developer studying how a small local vision model can drive a desktop, it is a clear, readable codebase. For getting work done, Interpreter Workstation and Eigent offer computer use inside a maintained product, and a 9B model will misclick on dense interfaces more often than the large hosted models those tools can call.
Where it earns its place
A machine learning engineer evaluating vision-language models
See how a quantised 9B model copes with real desktop tasks before choosing a model for a product
Serve the default Qwen checkpoint through vLLM at an 8192-token context, then swap the endpoint in settings to compare other models on the same tasks. Free software on hardware you already have; the 25-step cap keeps failed runs short.
A Linux desktop developer at a small open-source project
Prototype a local assistant that operates GUI apps without sending screenshots anywhere
Fork the repo and add tools or safety policies as TypeScript modules; keep Ask first enabled while testing so each click and keystroke needs approval. Run it under X11, where synthetic input is best supported.
Before you commit
- Setup
- Clone the repo, npm install with Node 22.12 or newer, start a vLLM server with the bundled launcher or point settings at a hosted endpoint, then npm start. Ubuntu and Debian may need libxtst6; macOS needs Screen Recording and Accessibility permissions.
- Integrations
- Any OpenAI-compatible Chat Completions endpoint with image input and tool calls; vLLM for local serving. Tools, providers and safety policies are TypeScript modules.
- Data and privacy
- With a local model, screenshots, prompts and tool results stay on the machine. The agent reads the screen and controls the real mouse and keyboard; API keys are stored in the app-data folder with user-only permissions, not the OS keychain.
- Cost at scale
- No licence cost. A local run needs a GPU with about 16 GB of memory for the default model; a hosted endpoint bills per token including screenshots.
- Limits
- Early alpha with one commit and no releases; primary display only; Wayland, Windows and macOS experimental; Windows cannot control elevated windows; 25-step default limit.
What’s good
- Runs fully local with a 9B vision model on a 16 GB GPU
- Ask first mode approves every mouse, keyboard and URL action
- Clear architecture with a sandboxed renderer
- MIT licence, any OpenAI-compatible endpoint
What’s not
- One commit, no releases, 31 stars
- Build from source only; no installers
- Only Linux X11 well supported; primary display only
- API keys not stored in the OS keychain
Best for
Developers who want to study or extend a small, fully local computer-use agent on a Linux X11 machine with a 16 GB GPU.
OpenComputer alternatives
OpenComputer: common questions
- Is OpenComputer worth it?
- OpenComputer is a tidy, honestly documented experiment in fully local computer use, but one commit, no releases and 31 stars make it a project to read rather than one to trust with your desktop. We rate it 2.9 out of 5.
- How much does OpenComputer cost?
- Free and open source (MIT). No paid plan; you supply the model, locally on your own GPU or through any compatible hosted API.
- What are the downsides of OpenComputer?
- One commit, no releases, 31 stars; build from source only; no installers; only Linux X11 well supported; primary display only; aPI keys not stored in the OS keychain.
- Who is OpenComputer best for?
- Developers who want to study or extend a small, fully local computer-use agent on a Linux X11 machine with a 16 GB GPU.
- How much does OpenComputer cost for a team or in production?
- No licence cost. A local run needs a GPU with about 16 GB of memory for the default model; a hosted endpoint bills per token including screenshots.
- Is OpenComputer safe for company data?
- With a local model, screenshots, prompts and tool results stay on the machine. The agent reads the screen and controls the real mouse and keyboard; API keys are stored in the app-data folder with user-only permissions, not the OS keychain.
- What are the best OpenComputer alternatives?
- The closest alternatives we have reviewed are Interpreter Workstation, Eigent, Magentic-UI (MagenticLite) and Browser Use.
Answers checked Nov 14, 2026.