Skip to content
AgentAtlas

Sandbox modes in every AI coding agent: Claude Code, Cursor, Codex and Gemini CLI compared (2026)

How do I run an AI coding agent's commands in a sandbox, and how do I allow network access?

Claude Code: run /sandbox or set sandbox.enabled and list hosts in sandbox.network.allowedDomains. Cursor: configure .cursor/sandbox.json, where networkPolicy defaults to deny. Codex: pick read-only, workspace-write or danger-full-access with --sandbox or sandbox_mode. Gemini CLI: start with -s, set GEMINI_SANDBOX or tools.sandbox, and choose Docker, Podman, macOS Seatbelt, gVisor or LXC.

Quick answer: how to turn it on

  • Claude Code: off by default. Run /sandbox, or set "sandbox": {"enabled": true} in ~/.claude/settings.json or .claude/settings.json.
  • Cursor: sandbox.json in ~/.cursor/ (all workspaces) or <workspace>/.cursor/ (one workspace, higher priority). type is workspace_readwrite (default), workspace_readonly or insecure_none, which disables it.
  • Codex: codex --sandbox read-only | workspace-write | danger-full-access, or sandbox_mode in config.toml. workspace-write is the default low-friction mode for local work.
  • Gemini CLI: gemini -s (or --sandbox), GEMINI_SANDBOX=true|docker|podman|sandbox-exec|runsc|lxc, or tools.sandbox in settings.json. The flag wins over the variable, which wins over the settings file.

What each sandbox uses under the hood

  • Claude Code: Seatbelt on macOS, bubblewrap and socat on Linux and WSL2. On native Windows commands run unsandboxed.
  • Codex: Seatbelt on macOS, bubblewrap on Linux and WSL2 (the first bwrap on PATH), and a native Windows sandbox in PowerShell.
  • Gemini CLI: Docker or Podman containers (default image ghcr.io/google/gemini-cli:latest), sandbox-exec (Seatbelt) on macOS, runsc (gVisor) or experimental LXC on Linux. runsc is never auto detected.

Network access

  • Claude Code: commands reach the network only through a local proxy. Allowed domains start empty; add hosts to sandbox.network.allowedDomains, block with deniedDomains, or set strictAllowlist to refuse anything unlisted.
  • Cursor: networkPolicy.default is deny unless you change it. allow and deny take exact domains, wildcards such as *.example.com (which also match example.com) or CIDR ranges. Deny always beats allow, URL paths are ignored, and private ranges, loopback and the cloud metadata IP are blocked by default.
  • Gemini CLI on macOS: pick one of six Seatbelt profiles with SEATBELT_PROFILE (permissive, restrictive or strict, each in an -open and a -proxied variant). The default is permissive-open, which confines writes to the project directory.

Claude Code: .claude/settings.json

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

Cursor: .cursor/sandbox.json

{
  "networkPolicy": {
    "default": "deny",
    "allow": ["registry.npmjs.org", "*.githubusercontent.com"]
  }
}

Codex and Gemini CLI: one run

codex --sandbox workspace-write --ask-for-approval on-request
gemini -s -p "analyze the code structure"

Extra write paths and escape hatches

  • Claude Code: sandbox.filesystem.allowWrite adds paths. sandbox.excludedCommands (for example docker compose *) runs trusted tools outside the sandbox. allowUnsandboxedCommands: false turns off the unsandboxed retry.
  • Cursor: readBoundary "workspace" (Cursor 3.23 or later) asks before reading outside the workspace. disableTmpWrite: true removes write access to /tmp and temp directories.
  • Codex: sandbox_workspace_write.writable_roots adds folders in workspace-write mode. danger-full-access removes the sandbox entirely.
  • Gemini CLI: GEMINI_SANDBOX_IMAGE or tools.sandbox.image picks a custom container image.

Full sandbox guide for each agent

Sources

Reviewed Oct 18, 2026. Settings change often; the linked vendor docs are the source of truth.

Built an MCP server? Get it listed everywhere →

Shipping an Android app? Google Play needs 12 testers for 14 days first. Get them with Tester Exchange →