Claude Code can deny sandboxed commands instead of prompting for network access

Claude Code's sandbox normally prompts you when a sandboxed bash command tries to reach a host that isn't on its network allowlist. Claude Code 2.1.219 (July 24, 2026) adds a setting to skip the prompt and deny instead.

{
  "sandbox": {
    "network": {
      "allowedDomains": ["github.com", "*.npmjs.org"],
      "strictAllowlist": true
    }
  }
}

With sandbox.network.strictAllowlist set to true, a sandboxed command reaching for any host outside the allowlist is denied outright — no approval dialog. The allowlist it checks against is network.allowedDomains plus any domains granted through WebFetch(domain:...) allow rules, or only the managed-settings entries when network.allowManagedDomainsOnly is also set.

Two scope limits matter here:

• It only gates sandboxed commands. In-process tools like WebFetch aren't affected by this setting.
• It's only honored from user settings, managed settings, or a CLI --settings value — not from .claude/settings.json or .claude/settings.local.json. A project can't flip this on for itself; it has to come from outside the repo.

Default is false, and the setting requires Claude Code v2.1.219 or later.

What this means

This is aimed at unattended or CI-style runs where a sandbox network prompt has nobody to answer it — previously that either blocked or (depending on your approval settings) silently allowed. strictAllowlist gives those runs a hard deny instead. See the Claude Code sheet for the current sandbox and permission settings.