Agent tools - BoxLite Documentation

These are the things an agent does inside a box. Each page owns one capability, its parameters, and a runnable example. If the agent should live inside the box rather than call into it, see Agent in a box instead.

In this section

Run Python

CodeBox and stdout back — the shortest path from an LLM’s output to a result.

Run any command

Any shell command or executable, in any language, in a disposable microVM.

Pseudo-terminal

A real interactive terminal, like docker exec -it.

Browser automation

A real Chromium, Firefox, or WebKit, exposed over CDP.

Computer use

A full Linux desktop the agent can see and click.

GitHub operations

Clone, commit, push, and open pull requests from inside a sandbox.

Drive from an agent loop

One safe door: the model proposes, your loop executes.

MCP server

Expose sandboxed execution to any Model Context Protocol client.

Which page do I want?

You are asking	Page
Run code an LLM just wrote	code execution python
Run a build, a test suite, or any binary	code execution any language
Keep a shell session open across turns	pseudo terminal
Let the agent browse the web	browser automation
Let the agent use a GUI	computer use
Wire this into an existing tool-calling loop	drive from agent loop
Expose it to Claude Desktop or another MCP client	mcp server

​In this section