What is open-cowork?

open-cowork is an open-source tool that lets you describe a computer task in plain text and then watch an AI agent carry it out on screen, step by step. You can point it at your own desktop, a cloud virtual machine, or a browser window. As the agent works, you see a live stream of what it is doing, and it pauses to ask for your approval before taking actions that matter. The project is described as an open-source alternative to Claude Co-Work, built using the Coasty Computer Use API by default. That API handles the part where an AI model looks at a screen and decides what to click or type next. If you do not want to use Coasty, you can switch to any other provider in settings: OpenRouter, OpenAI, or a locally running model through Ollama or LM Studio. This is the BYOK (bring your own key, or bring your own LLM) mode. Setup takes one command after cloning the repo: pnpm desktop. That starts the backend, opens the web interface, and launches a desktop app that can control your actual mouse and keyboard. With no API key configured it runs in demo mode using a mock server and a sandbox key, so nothing is billed and no account is needed to try it out. For more involved use cases, the tool includes a workflow builder where you write tasks as a JSON description that supports branching, loops, retries, parallel steps, and human approval checkpoints. Each run shows a cost estimate upfront and requires you to confirm before anything billable starts. There are also hard budget caps enforced on the server side. You can monitor and manage runs from a web dashboard, a desktop window, or a mobile app. If a run pauses waiting for your approval, a notification appears on your phone so you can unblock it from wherever you are. The event log is durable and can replay from the beginning, which helps when you want to review what the agent did. The codebase is a TypeScript monorepo (Node 22+, pnpm) covering the backend API, the web frontend, the Electron desktop app, and a React Native mobile app.