Anthropic launches Claude Code web interface with autonomous sandboxing

Anthropic has launched web and mobile interfaces for Claude Code, its popular command-line AI coding tool, while introducing a new sandboxing system that allows the agent to operate more independently. The web version enables developers to connect GitHub repositories and give high-level instructions like “add real-time inventory tracking to the dashboard,” while the enhanced sandboxing reduces the need for constant approvals by pre-authorizing specific folders and network access.

What you should know: The web interface brings Claude Code’s capabilities to browsers with full GitHub integration and multi-session support.
• Developers can run multiple coding sessions simultaneously and switch between them using a left-side panel interface.
• The tool now accepts mid-task suggestions and changes, eliminating the need to cancel and restart when corrections are needed.
• A limited iOS mobile version is also available, though it’s in earlier development stages.

The big picture: The new sandboxing runtime represents a fundamental shift from permission-heavy interactions to pre-authorized, secure automation.
• Previously, Claude Code required approval for most changes and steps throughout the coding process.
• The updated system allows users to grant permissions for specific file system folders and network servers upfront.
• Network isolation restricts internet access through a specialized proxy server running outside the sandbox, with domain-level restrictions and user confirmation for new requests.

In plain English: Think of the old system like having to ask permission every time you want to open a door in your house. The new approach is like giving someone a key to specific rooms and telling them which websites they can visit—more freedom within safe boundaries.

Why this matters: These changes make AI coding agents more autonomous while introducing new risks that require careful consideration.
• The reduced friction allows Claude Code to work more independently without constant line-by-line approvals.
• However, fewer approval steps mean developers might more easily miss when the AI makes poor decisions.
• Code review becomes even more critical as the tool operates with greater independence.

How it works: The sandboxing system balances automation with security through customizable proxy controls.
• Users can customize the proxy to set their own rules for outgoing network traffic.
• The system allows fetching software packages from approved sources without giving “carte blanche for communicating with the outside world.”
• This approach reduces constant approval requests while maintaining security against prompt injection and other risks.

What’s available: The new features are currently in beta as a research preview for Claude Pro and Max subscribers.