OpenBot gives AI agents their own isolated sandboxes

Handing an AI agent access to your browser is terrifying—unless it gets its own isolated sandbox.

CopilotKit launched OpenBot, an open-source runtime that equips AI agents with their own sandboxed browsers, filesystems, and shells. The project gained 1,578 stars on GitHub this week. It connects to agents built on LangGraph, CrewAI, or Pydantic AI using the open AG-UI protocol, routing every tool call through a central gateway that enforces rules and records audit logs.

Why it matters: Most frameworks let models run commands directly on your machine without guardrails. OpenBot puts each agent inside its own container with a dedicated Chromium instance and workspace. You can watch the agent work live, set execution boundaries in the admin panel, and take manual control if it reaches a sensitive step.

It runs locally on your machine using Docker and Bun, though keep in mind it is currently in alpha. You can run the stack with scripts/start.sh, set up deny rules at /admin/boundaries, or add custom coworkers by editing agents.yaml.

Giving an AI its own sandbox beats giving it your local permissions every single time.