Should MCP servers start with zero network? A sandboxed Claude Code setup argues yes
quiet_flare · reddit · 2026-11-01
A developer running Claude Code inside an NVIDIA OpenShell sandbox with just two network-policy rules (each naming a host and one allowed binary) argues MCP servers should start with no network at all. Telemetry pings and update checks at startup either fail or force widening the allowlist. His proposal: tools should connect to remote APIs only when invoked, so policy gates can inspect and refuse requests, READMEs should declare which hosts each tool reaches, and the community needs a machine-readable per-tool egress manifest convention. He shares the setup as openshell-claude and asks what current egress-declaration practices exist.
More from coding & agent
- Reddit user details Codex disaster: broken UI and botched logic in Unity game port — AcademicMistake · 2026-11-01
- 'The Secret Is Error-Decoupled Agentic Scaling' — and It's Not on Your API Yet — tokenbender · 2026-11-01
- Parkinson's Law for agents: more budget just means more tokens, not more work — tokenbender · 2026-11-01
- PocketShell turns your Android phone into an AI Linux workstation for Claude Code agents — pocketshelldev · 2026-11-01
- Two crashed runs led to 6 rules: repair your agent's plan file at load time — petrusenko_max · 2026-11-01
- Six human-in-the-loop patterns for AI agents, from log-only to two-person approval — Relevant_Strategy_38 · 2026-11-01