Anthropic open-sources sandbox-runtime, the OS-level sandbox behind Claude Code
solyarisoftware · x · 2026-10-24
Anthropic has open-sourced sandbox-runtime, the lightweight sandboxing tool used by Claude Code. It enforces filesystem and network restrictions at the OS level on arbitrary processes — using macOS's built-in Seatbelt and bubblewrap on Linux — without requiring a full container. It works not only with Claude Code but also with your own agents, MCP servers, or plain CLI tools.
The post also flags the real challenge: agents need to install dependencies, fetch docs, and call APIs. Whitelists that are too tight break tasks; too loose, the sandbox loses its value. The hard part of agent sandboxing is balancing security with usability.
More from coding & agent
- When Should an Agent Stop Trying? A Developer's Case for Explicit Stop Criteria — artemgetman · 2026-10-24
- RLCR trains LLMs as agent orgs with shared workspaces to automate companies and markets — rosstaylor90 · 2026-10-24
- Plane explains agent access control: scope comes from three independent layers, not project limits — JosephJacks_ · 2026-10-24
- Do Agents Actually Learn? Persistent Memory Experiments Show Mixed Results — artemgetman · 2026-10-24
- Cutting KServe cold starts from 134s to 9s with warmctl — kbesplv · 2026-10-24
- Multi-User MCP Server With OAuth: 89 Broken Versions Tested, 32s Cold Start Workarounds — ShipItFrog · 2026-10-24