When you talk to Claude 20 times a day across a dozen different projects, one screen stops being enough. Your agent triggers a file download. Your other agent tries to upload something. One runs a test that crashes X. Another needs a screenshot at the exact moment the first one refreshes the page. Chaos.
That's the problem Screenbox solves: each AI agent gets its own isolated Linux desktop. Same hardware. Complete separation. I can run 5, 10, or 20 agents in parallel, and they don't step on each other.
Let me be clear about what this isn't. Screenbox is not a browser sandbox -- it's a full desktop for agents, with real Chromium, a file system, terminal access, and 100+ actions. It's not SaaS cloud infrastructure -- it's self-hosted, AGPL-3.0, runs on your own hardware. It's not designed to replace your existing agent framework; it's designed to give agents a controlled workspace to actually do things without breaking each other or your sanity.
How I got here
For most of 2025, I was building Typus -- a framework for launching services fast. By the time it shipped, AI had moved so fast that the problem it solved didn't matter anymore. But something shifted for me: instead of building frameworks, I started building products by talking. Voice input to Claude. Screenshots to see what happened. Commands to execute. It was faster.
At first, one desktop worked. But by month three, I had dozens of projects running in parallel. My workstation looked like a meth lab -- windows scattered everywhere, agents fighting for resources, one bad interaction pulling down the whole session. I couldn't tell what was happening. I couldn't scale.
So I built Screenbox.
What you get
Each agent runs in a Docker container with its own isolated desktop environment. Full Linux, real Chromium, file system. When an agent crashes, it crashes alone. When it downloads a file, that file goes to its own directory. When it manipulates the DOM, the DOM is its own to break.
Screenbox exposes 21 MCP tools -- the standard protocol for connecting Claude to external systems. The agent can take screenshots, click elements, type text, run shell commands, manipulate windows, drag files. Everything a human sitting at a Linux desktop can do.
I've been running this in production for 4+ months. The metrics look boring, which is exactly what you want: 847 Chrome restarts across all agent sessions. That number doesn't mean Chrome crashes -- it means agents are running long enough that Chrome occasionally restarts for memory management or to clear stale state. Under real load, from agents running actual complex tasks. No panics. No silent failures. Just steady, reliable isolation.
The real problem it solves: resource contention
Last year, I gave two agents the same desktop and asked them to work on different files. Agent A downloaded a spreadsheet to /tmp/workspace/file.csv. Agent B needed to download its own file to the same location. Collision. File was corrupted. Agent B got garbage. I had no visibility into what happened until it was too late.
Another time: one agent ran a CPU-intensive task on the shared X server. The other agent's screenshot failed. Its clicks didn't register. I debugged for 20 minutes before realizing the problem wasn't in the code -- one agent was consuming all available GPU/CPU resources. With isolated desktops, that agent gets its own resources, and the contention just... disappears.
That's the insight: isolation is safety. You get throughput, reliability, and visibility without complex orchestration or resource quotas.
What I didn't expect: agents learning
Screenbox has a knowledge compilation system. When an agent completes a task and encounters similar tasks again in future sessions, it gets faster. It remembers what worked. The learning is per-agent, saved to its local context. Over time, repeated tasks get 10-20% faster as the agent narrows down what needs to happen.
I didn't design this as a feature. It just emerged from isolation -- each agent builds its own knowledge map without interference. Now I'm leaning into it.
Who needs this
If you're running more than one AI agent, and they're doing real work -- browsing, downloading files, running tests, manipulating the filesystem -- you need isolation. Right now, if you're using Computer Use API or trying to run agents locally, they're sharing your system resources. Screenbox gives each one its own box.
How to try it
Screenbox is open source (AGPL-3.0), self-hosted. No signup. No cloud. Run it on your server.
GitHub: github.com/dklymentiev/screenbox Docs & examples: screenbox.dev
There's a getting-started guide that takes 5 minutes. Docker Compose setup. One agent session to see how it works. From there, you'll either need this or you won't.
The mental model
I built a room for them to work in. Walls so they couldn't knock things over. Each one has its own desk, its own files, its own screen. A window so I can watch what they're doing. When one makes a mess, the others keep working. When I need to see what happened, I just look.
That's Screenbox.