drun offers an open-source safe workspace for AI agents
drun is an initial MCP release for giving AI agents a safer workspace. It creates an by virtualizing parts of the host computer, so the agent can work in a separated environment. The agent can use Git-like controls to try several paths in parallel and throw away paths that fail.
The main goal is to avoid changing or damaging the host state while the agent explores options. drun adds hard rules for network access, command execution, paths, memory use, and run time. Instead of giving an agent direct create, read, update, and delete access on the host, drun puts a customizable policy layer in between.
The project is open for , feedback, and community contributions.
Key points
- drun is an MCP for safer AI agent workspaces.
- It gives agents an separated from the host computer.
- Agents can try multiple paths and discard failed paths without changing the host state.
- It can restrict network domains, commands, access, memory, and duration.
- Its value is more about safe agent building and operating-cost control than direct token reduction.