You'll noticeTier: how much it should matter to you
2Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
SandboxArea: what it touches
ImprovementsKind: in v2.1.299,
ImprovementsSection of the release
What
The sandbox is a protected space where Claude Code runs commands with limited access to your system. The instructions Claude Code gives the model about the sandbox now include Windows sections:
Sandboxed commands cannot connect to localhost (127.0.0.1 or ::1), even to their own port, and no setting allows it
Notes on Maven and .NET, pip certificate errors, and the 0-byte placeholder files the sandbox creates, such as .mcp.json and .claude/settings.json
The suggested fixes, running /sandbox exclude or prefixing a command with !, are now offered only when they apply, for example not when your settings are locked by an organisation policy.
When Windows refuses to start Claude Code as the sandbox account with an access-denied error, Claude Code now tells the model to retry once and resets its permission to restart itself in the sandbox.
Why
On Windows, the model can now explain more accurately why a sandboxed command failed and suggest a fix that actually works in your setup.
How sure we are
One source agreesOne thing we can check says the same as this entry.
Anthropic's release notes agreeFixed claude mcp list, claude doctor and /status reporting an empty .mcp.json as invalid
