Screenbox
screenbox.dev
- Free plan
- Windows
- macOS
- Linux
- Web
- API
- Self-hosted
What Screenbox is
Screenbox provides self-hosted virtual desktops for AI agents, with real browsers and desktop applications. Each desktop includes Chromium with DevTools and extensions, along with tools such as file managers, terminals, and office applications. Agents can use 21 MCP tools across more than 100 actions, including screenshots, OCR, clicking, typing, shell, file, and clipboard operations. It connects to MCP clients such as Claude Desktop, Claude Code, and Cursor through stdio, streamable HTTP, or SSE. A web dashboard provides live viewing, agent management, snapshots, session logs, and a knowledge browser. Users can save and restore desktop state, including files and browser sessions, or clone desktops from templates. Desktops run in Docker containers with memory limits, network isolation, and no bind mounts; the maker cautions these are not hardened sandboxes. Screenbox is free, open-source, and runs on the user's infrastructure. The maker says self-hosted use sends no telemetry or data to Screenbox or third parties. Desktop use needs about 2 GB RAM and no GPU.
Who it suits
Screenbox suits teams building AI-agent workflows that need browser and desktop interaction in a self-hosted environment. It also fits users who need to observe or take control of agent activity.
Good to know, and what to watch
- 21 MCP tools span more than 100 actions.
- Save and restore desktop state and browser sessions.
- Users can watch and control agent activity.
- Self-hosted deployment sends no telemetry, according to the maker.
- Docker containers are not hardened sandboxes.
- Each desktop uses about 2 GB RAM.
- The maker advises against exposing the MCP API publicly.
Plans
Every plan and priceSelf-hosted
Open source
Getting started
- Visit https://screenbox.dev/.
- Choose the self-hosted plan, which is free under AGPL-3.0.
- Prepare Linux with native Docker, macOS with Docker Desktop, or Windows with WSL2.
- Run setup to generate API tokens automatically.
- Connect an MCP client using stdio, streamable HTTP, or SSE.
- Use localhost or VPN access for the MCP API, as advised by the maker.
Limits to check
The self-hosted plan runs on your own infrastructure. Each desktop uses about 2 GB of RAM, and the Docker containers are not hardened sandboxes.
The Club's review
Screenbox is an open-source, self-hosted environment where AI agents operate real browsers and desktop applications inside disposable virtual desktops. It suits developers and IT teams building computer-use workflows that need supervision, repeatable state, and local data control. The main trade-off is that running Docker containers still requires careful security design.
Overview
Screenbox places each agent in a Docker-based desktop with Chromium, desktop utilities, and access to the screen, files, shell, and clipboard. Connections work through Claude Desktop, Claude Code, and Cursor using stdio, streamable HTTP, or SSE transports, so it can fit both local development and service-style integrations.
The web dashboard combines live desktop viewing, agent management, snapshots, session logs, and a knowledge browser. Users can watch through RDP or noVNC and take mouse and keyboard control when an agent needs intervention. That combination is valuable for workflows that cannot be left entirely unattended, while self-hosting leaves deployment, updates, and access control with the operator.
Key features
- Broad MCP automation: Twenty-one MCP tools expose more than 100 actions, including screenshots, OCR, clicking, typing, shell commands, file I/O, and clipboard operations. This is a better fit for cross-application tasks than systems limited to browser actions.
- Real browser and applications: Every desktop includes Chromium with DevTools and extensions, plus applications such as file managers, terminals, and office tools. Agents can therefore work with desktop software as well as web pages.
- Human control: RDP and noVNC viewing, with mouse and keyboard takeover, gives operators a practical recovery path when an agent stalls or reaches an ambiguous step.
- Snapshots and templates: Desktop state can be saved, restored, and cloned, including files and browser sessions. This supports repeatable runs and quicker recovery from failed automation.
- Reusable knowledge: Session action logs can be compiled into facts and injected into later screenshot and look responses, helping recurring workflows retain operational context.
- Resource controls: Idle desktops pause automatically, and a configurable time-to-live can release acquired desktops. Each desktop uses about 2 GB of RAM and does not need a GPU.
- Isolation with a clear boundary: Containers have memory limits, network isolation, and no bind mounts. The maker nevertheless warns that containers are not hardened sandboxes, so sensitive or hostile workloads need stronger separation and the MCP API should stay on localhost or behind a VPN rather than being exposed publicly.
- Privacy: The self-hosted software sends no telemetry or data to Screenbox or third parties, keeping deployment data on the operator's infrastructure.
Pricing
Self-hosted — free. This is AGPL-3.0 open-source software running on your own infrastructure. There are no paid seats or hosted usage tiers described, but you supply the machines, Docker environment, maintenance, and operational security yourself. It is the natural choice for teams that need local deployment and API access; readers wanting a managed service should look elsewhere.
Platforms
Screenbox supports Linux, macOS, Windows, web, API, and self-hosted deployment. The documented setup uses native Docker on Linux, Docker Desktop on macOS, and WSL2 on Windows. Desktop control, browser control, and screen understanding are supported, with local deployment and API access. The roughly 2 GB RAM requirement per desktop makes capacity planning important when running several agents at once.
Who it's for
Screenbox is aimed at builders of AI computer-use agents, automation teams, and organizations that need to observe and intervene in desktop workflows while keeping data on their own infrastructure. Snapshots and reusable session knowledge suit repeatable processes. It is a poor fit for teams that cannot operate Docker, configure network access, or accept that container isolation is not a hardened security boundary.
Pros and cons
Pros
- Open-source and free to run, with deployment data kept on your infrastructure.
- More than 100 desktop and browser actions through 21 MCP tools.
- Real Chromium, extensions, DevTools, and desktop applications support broader workflows than browser-only agents.
- Live oversight, takeover controls, snapshots, cloning, and session knowledge support recoverable automation.
- Automatic pausing and configurable release times help limit idle resource use.
Cons
- Docker isolation is explicitly not a hardened sandbox, limiting suitability for untrusted workloads.
- Self-hosting shifts infrastructure, upgrades, capacity, and security configuration to the user.
- Exposing the MCP API publicly is discouraged, so remote access requires a localhost or VPN-oriented design.
- About 2 GB of RAM per desktop can become a significant requirement for parallel agents.
Alternatives
Choose AI Computer Use Agents for a broader category comparison. Claude Code is a paid, multi-platform option when you want a general agent product rather than Screenbox's self-hosted desktop infrastructure. Verdesk suits Windows users seeking a freemium browser-focused tool. Computer Use OOTB is another free, open-source choice for locally running or API-based models on Windows and macOS. Gemini is a freemium assistant spanning web, desktop, and mobile. Lapu AI offers free and paid agent plans on macOS and Windows. agent-cu is a locally running MIT-licensed alternative with zero per-action cost. ChatGPT for Excel is the focused choice for spreadsheet work in a web interface. Scalaris Atlas is a free, cross-platform option with zero egress by default.
Verdict
Choose Screenbox if you are building or operating AI agents that need real desktop applications, human takeover, recoverable state, and self-hosted data control at no software cost. Its MCP breadth and snapshot workflow are substantial advantages for supervised automation. Look elsewhere if you require a managed service or a security boundary stronger than Docker containers.
The verdict
Screenbox combines agent-accessible desktops with snapshots, monitoring, and human control. Its Docker isolation has limits, so the maker's warning against treating containers as hardened sandboxes matters.
Questions people ask
How much does Screenbox cost?
The Self-hosted plan is free and is AGPL-3.0 open-source software that runs on your own infrastructure.
Which operating systems does Screenbox support?
The maker lists Linux with native Docker, macOS with Docker Desktop, and Windows with WSL2.
Which MCP clients can connect to Screenbox?
It connects to clients including Claude Desktop, Claude Code, and Cursor, using stdio, streamable HTTP, or SSE transports.
Is Screenbox a hardened sandbox?
No. Desktops run in Docker containers with memory limits, network isolation, and no bind mounts, but the maker cautions that the containers are not hardened sandboxes.
Does Screenbox send telemetry?
The maker says the self-hosted software sends no telemetry or data to Screenbox or third parties, and deployment data remains on the user's infrastructure.
What the maker publishes
- Automation
- It offers 21 MCP tools across more than 100 actions, including screenshots, OCR, clicking, typing, shell, file I/O, and clipboard operations. source
- Browser and apps
- Each desktop includes real Chromium with DevTools and extensions, plus desktop applications such as file managers, terminals, and office tools. source
- Dashboard
- The web dashboard provides live desktop viewing, agent management, snapshot controls, a knowledge browser, and session logs. source
- Human oversight
- Users can watch agent activity through RDP or noVNC and take mouse and keyboard control. source
- Integrations
- It connects to MCP clients including Claude Desktop, Claude Code, and Cursor, and supports stdio, streamable HTTP, and SSE transports. source
- Isolation
- Desktops run in Docker containers with memory limits, network isolation, and no bind mounts; the maker cautions that containers are not hardened sandboxes. source
- Knowledge
- Screenbox can compile session action logs into reusable facts and inject that knowledge into future screenshot and look responses. source
- Maker
- The Terms identify Dmytro Klymentiev as the product's developer and maintainer. source
- Platforms and requirements
- The maker lists Linux with native Docker, macOS with Docker Desktop, and Windows with WSL2; each desktop uses about 2 GB RAM and needs no GPU. source
- Privacy
- The maker says the self-hosted software sends no telemetry or data to Screenbox or third parties, and deployment data remains on the user's infrastructure. source
- Purpose
- Screenbox provides isolated virtual desktops with real browsers and desktop applications for AI agents. source
- Resource management
- Idle desktops pause automatically, and acquired desktops can be released after a configurable time-to-live. source
- Security configuration
- The maker advises against exposing the MCP API publicly, recommends localhost or VPN access, and says API tokens are generated automatically by setup. source
- Snapshots
- Users can save and restore desktop state, including files and browser sessions, and clone desktops from templates. source
- Support
- The security policy asks vulnerability reports to be submitted through GitHub private vulnerability reporting and says it aims to acknowledge them within 48 hours. source
Checked against the maker's pages on 7 Oct 2026.