Anthropic has open-sourced its sandbox-runtime (srt), an OS-level tool that provides sandboxing for AI agents. The runtime limits file access and network connectivity for processes without requiring containers. It is compatible with custom agents, local MCP servers, and plain command-line programs.
Key points
- Anthropic's sandbox-runtime (srt) is now open-source.
- It imposes file and network restrictions on AI agents.
- The tool operates without using containers.
- Internal testing showed an 84% reduction in permission prompts.
The npm package, @anthropic-ai/sandbox-runtime v0.0.81, is approximately 11.4 MB unpacked and released under the Apache-2.0 license. It requires Node.js 22.12+ and runs on macOS and Linux development machines or CI systems. Windows support is currently in an alpha stage.
On macOS, srt utilizes Seatbelt for isolation, while Linux requires bubblewrap, socat, and ripgrep. Anthropic reports that internal usage of srt reduced permission prompts for Claude Code by 84%, enforcing file and network limits across the entire process tree.
By the numbers
- Reduction in permission prompts
- 84%
- Unpacked npm package size
- 11.4 MB
The sandbox-runtime functions as a command-line interface (CLI) and a TypeScript library. Users can install it via npm and prefix commands with 'srt' to apply restrictions. The tool leverages each operating system's native isolation primitives to enforce its security policies on the entire process tree.
Sources and references
Original report
Meet Anthropic’s sandbox-runtime: OS-Level Sandboxing for AI Agents, No Containers
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Codegithub.com/anthropics/sandbox-runtime (opens in a new tab)
- Linknpmjs.com/package/@anthropic-ai/sandbox-runtime (opens in a new tab)
- Codegithub.com/containers/bubblewrap (opens in a new tab)
- Linkanthropic.com/engineering/claude-code-sandboxing (opens in a new tab)
- Linklearn.chatgpt.com/docs/sandboxing (opens in a new tab)
- Linkdocs.docker.com/ai/sandboxes/security/policy (opens in a new tab)
Ref 001CI6HU
Cite this
APA
Interestana AI Editorial. (2026, October 24). Anthropic Releases Open-Source AI Agent Sandbox Runtime. Interestana. https://interestana.com/articles/meet-anthropics-sandbox-runtime-os-level-sandboxing-for-ai-agents-no-containers-001ci6hu
MLA
“Anthropic Releases Open-Source AI Agent Sandbox Runtime.” Interestana, 24 Oct. 2026, interestana.com/articles/meet-anthropics-sandbox-runtime-os-level-sandboxing-for-ai-agents-no-containers-001ci6hu.
Chicago
Interestana AI Editorial. “Anthropic Releases Open-Source AI Agent Sandbox Runtime.” Interestana, October 24, 2026. https://interestana.com/articles/meet-anthropics-sandbox-runtime-os-level-sandboxing-for-ai-agents-no-containers-001ci6hu.
BibTeX
@misc{interestana_001ci6hu,
author = {{Interestana AI Editorial}},
title = {{Anthropic Releases Open-Source AI Agent Sandbox Runtime}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/meet-anthropics-sandbox-runtime-os-level-sandboxing-for-ai-agents-no-containers-001ci6hu},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/24/anthropic-sandbox-runtime-ai-agent-sandbox/}
}Credit the source
Original reporting: MarkTechPost, “Meet Anthropic’s sandbox-runtime: OS-Level Sandboxing for AI Agents, No Containers”, https://www.marktechpost.com/2026/10/24/anthropic-sandbox-runtime-ai-agent-sandbox/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe