Show HN: Screenbox – Self-hosted virtual desktops for AI agents
TL;DR
Screenbox provides each AI agent its own isolated Linux desktop environment with a real Chromium browser, controlled via MCP (Model Context Protocol).
Key Points
- Each environment runs as a Docker container using around 2 GB RAM, no GPU required.
- Multiple agents can run in parallel without conflicting – solving the exact problem that inspired the project.
- Users can watch agents work in real time and take over control whenever needed.
- Open source under AGPL-3.0, self-hosted at screenbox.dev.
Nauti's Take
Classic 'scratching your own itch' tooling – and that is exactly why it is convincing. No bloated feature set, just a direct answer to a real problem in parallel agent workflows.
Two gigabytes of RAM per desktop is acceptable as long as tasks are not overly resource-intensive. The interesting question is how deeply Screenbox can integrate with frameworks like LangGraph or AutoGen going forward.
The AGPL license is a deliberate move to prevent commercial players from simply forking and monetizing the work.
Context
Anyone running multiple AI agents on browser-based tasks simultaneously knows the problem: shared browser sessions create chaos. Screenbox solves this structurally by bringing browser isolation to the container level – without cloud dependency. For developers testing or running agent pipelines locally, a lightweight self-hosted setup is far more attractive than hosted alternatives like Browserbase.
The MCP integration makes it directly compatible with modern agent frameworks.
Sources