Anthropic Sandbox Runtime: limit what AI agents can touch on your computer
Published· By Mishal Brands Editorial · ai-news, ai-tools
Anthropic's open-source Sandbox Runtime wraps a command or AI agent in OS-level file and network limits, without a container. Here is what it does and how to try it.
Anthropic Sandbox Runtime (srt) is an open-source tool that puts a command inside file and network limits (GitHub). Anthropic describes it as a research preview developed for Claude Code.
What it does, as reported
- Uses your OS sandbox. sandbox-exec on macOS, bubblewrap on Linux, and an alpha Windows mode, plus proxy-based network filtering.
- Locked down by default. Network access is denied and writes are allowed only where you list them.
- Works with MCP servers. The README shows putting srt in front of a local MCP server command.
- CLI or library. Use it from the terminal or inside your own code.
Why it matters
AI agents can run commands. A sandbox means a mistake, or a bad instruction hidden in a web page, can't reach your whole computer.
How to try it
- Install it:
npm install -g @anthropic-ai/sandbox-runtime(npm). - Run a command inside it, following the README's examples.
- Add only the folders and domains the task needs.
It is free and open source under the Apache-2.0 license, per its npm package.
Tools in this post
- Anthropic Sandbox Runtime: Open-source sandbox that limits the files and network an AI agent or command can reach.