Running Codex, Claude and Pi Agents Safely: gVisor Sandboxes Plus tart macOS VMs
craigbalding · x · 2026-09-18
A detailed first-hand setup for running coding agents (codex, claude, pi) safely: all harnesses run in gVisor-protected microvms where agents get root and can install anything, yet stay sandboxed. For macOS work, tart on the same Mac mini runs a Tahoe VM, accessed by the agent via a shell bridge that executes commands with tart exec. Separate seatbelt-restricted SSH access enables first-class vz testing and structured MLX experiments with no network in/out — files are downloaded via a controlled proxy in the container and streamed over SSH via HTTP CONNECT.
Related event: Running Codex/Claude Agents Safely in microVM+gVisor Sandboxes(2 posts)→
More from coding & agent
- Pydantic AI agents now run on TypeSafe's Jev with per-field confidence outputs — Paimaamu · 2026-09-18
- Demo: GPT-6 Astra edits video in DaVinci via MCP at impressive speed — toolstelegraph · 2026-09-18
- After a day with Jev: a blazing-fast classifier, not a GPT replacement — jiayuan_jy · 2026-09-18
- Wiring Claude Code, Codex and Cursor to one shared MCP memory server — Asly97 · 2026-09-18
- Running Claude Code, Codex and Cursor on One Project: How I Finally Stopped Losing Context — Asly97 · 2026-09-18
- Multica hits 50k GitHub stars, ships cloud agent integration with Qoder — jiayuan_jy · 2026-09-18