Agent-Infra unveils AIO Sandbox, a single environment for AI agents with a browser and shell

Agent-Infra has released AIO Sandbox, an open-source runtime for AI agents that packs a browser, terminal, file system and MCP services into one container. The project does not aim to improve the reasoning model. It targets a more practical problem: giving an agent an isolated, convenient place to take real actions.

The agent bottleneck

Many agent scenarios are now limited not by the model's intelligence but by the infrastructure around it. An LLM can write code, plan steps and call tools, but then the chaos starts: the browser lives in one place, the Python runtime in another, the shell in a third, and files have to be passed between them by hand. Each hand-off adds latency, breaks session state and requires extra glue code that is hard to maintain.

On paper, Docker containers solve this, but real pipelines run into a different problem: fragmentation. If an agent downloads a CSV through the browser, the file still has to be moved into the processing environment, the result saved, and then handed to the next tool. The more steps, the higher the risk of desynchronization and errors at the seams between services. AIO Sandbox tries to remove exactly this layer of engineering routine and give the agent one workspace instead of a set of loosely connected modules.

What's inside the Sandbox

AIO Sandbox is essentially a single Docker runtime that already bundles the main interfaces a modern agent needs. Instead of configuring browser automation, a terminal, an IDE and a connection protocol separately, developers get a ready-made environment with native access to all of these components. The repository describes the project as an all-in-one sandbox environment aimed both at agents and at developers who want to observe, debug and reproduce system behavior.

• Chromium controlled via CDP, with Playwright support
• Preconfigured Python and Node.js runtimes
• Bash terminal and a shared file layer for all modules
• Built-in VSCode Server and Jupyter Notebook for debugging
• MCP servers for the browser, files, shell and document-to-Markdown conversion

The key idea is a unified file system. A file the agent downloads through Chromium is immediately available to a Python script, a bash command or the file API, with no intermediate transfers. For agent tasks this matters more than it sounds: an agent can open a web portal, export data, clean it in Python right away and save the result in the same working directory.

The MCP integration is also useful: the tools are already prepared and do not require the team to write manual wiring for basic operations.

How it runs

Agent-Infra stresses that the Sandbox suits not only local experiments but also production deployment. The documentation offers a quick start with a single Docker command, as well as examples for Docker Compose and Kubernetes. The container can be limited by CPU and memory, and agent code runs isolated from the host system.

The environment also stays stateful: terminal sessions and working state can be preserved across a multi-step task rather than reset after every action.

On top of the runtime, an API and SDK let an application read and write files, run shell commands, control the browser, execute Jupyter code and connect MCP services. A VNC interface for visually monitoring the browser and VSCode Server for manually checking what the agent is doing are also available out of the box.

This combination is especially useful for teams building web operators, internal automation workflows or hybrid pipelines that need to combine site navigation, local data processing and verifiable isolation.

What it means

If the trend toward agent systems continues, the next area of competition will be not only model quality but also the quality of the execution environment. AIO Sandbox shows where the market is heading: toward a standard runtime in which the browser, code, files and tool protocols are already wired together. For teams, it is a chance to spend less time on infrastructure plumbing and get agents from demo to working service faster.