Anthropic makes Claude Code accessible via the browser and introduces a new sandboxing feature designed to enhance developer safety and autonomy.
Claude owner Anthropic announces that the coding tool Claude Code is now also available via the web. Claude can be accessed directly via the browser, without additional tools or terminals. Simultaneously, a mobile version is also being released, although it is limited to iOS at launch.
read also
Anthropic Bridges the Gap between Claude and Microsoft 365
The new web interface allows for multiple coding tasks to be executed simultaneously on Anthropic’s cloud infrastructure, hosted on AWS or Google Cloud. Developers can link GitHub repositories, describe tasks, and Claude handles the implementation. Each session runs in a separate, isolated environment where progress can be monitored in real-time. Claude can also be adjusted during the process. The tool automates pull requests and provides an overview of the changes made.
This approach is intended for routine tasks such as bug fixes, test-driven backend adjustments, or documentation queries, according to Anthropic. The web interface also works on mobile via Claude’s iOS app; an expansion to Android is still pending. The current version is a research preview and is available to Pro and Max users.
Sandbox
Anthropic also introduces a new sandboxing mechanism that gives Claude more freedom within clearly defined boundaries. This mechanism significantly reduces the number of approval prompts and lessens the risk of security issues, such as prompt injection. In internal tests, this led to 84 percent fewer prompts.
The sandbox consists of two layers of protection: file isolation and network isolation. Claude can only modify files within predefined folders and connects to approved network locations via a proxy server. The proxy verifies Git actions, manages access rights, and prevents unauthorized interactions with external systems. This helps protect against the leakage of sensitive data such as SSH keys.
Additionally, a new bash tool is available within Claude Code, built at the operating system level. This tool allows for the execution of secure bash commands without constant user approval. If an attempt is made to access outside the sandbox, the user receives a notification. Both features are available open source so that other development teams can also integrate them.