• Dotika
  • Posts
  • Claude Code gets web version

Claude Code gets web version

ALSO : OpenAI launches AI browser for Mac

News of the day

1. Anthropic launches web and mobile interfaces for its AI coding tool, Claude Code. Key improvements focus on mid-task adjustments and enhanced sandboxing. Read more

2. OpenAI launches ChatGPT Atlas, an AI-integrated browser for Mac, blending ChatGPT's capabilities into web browsing for enhanced productivity and automation.  Read more

3. EU data regulations like GDPR are reducing AI innovation, with cultural factors influencing the impact. Research suggests a trade-off between privacy and competitiveness.  Read more

4. China's generative AI user base has doubled to 515 million in six months, driven by infrastructure and state support, with a strong preference for domestic models.  Read more

Our take

Hi Dotikers!

Anthropic extends Claude Code to the browser. Many will focus on the interface and convenience, but here’s what really matters: you connect your GitHub repositories, launch multiple tasks in parallel, and track their progress in real time on Anthropic’s managed infrastructure. Each session is isolated. This is an early preview for Pro and Max users, with an iOS glimpse as well. The result: cloud-based sessions, automatic PRs, and clear change summaries.

At the core of this update is the new sandboxing model. Anthropic introduces two simple, intuitive barriers: one for files and one for the network. Git operations go through a proxy that only exposes credentials for authorized repositories and branches. The goal is to let the agent work more freely within a defined perimeter ; without triggering a permission alert for every single command.

As for measurable impact, Anthropic reports an 84% reduction in permission requests in internal use. Under the hood, the isolation relies on system-level components: bubblewrap on Linux and seatbelt on macOS. To help the broader ecosystem align, the company is also open-sourcing its sandbox runtime, available as both a CLI and a library to enforce the same file and network rules.

The web interface is a nice addition, but the real value lies in the containment model. The tools that will win out aren’t those that code the fastest—they’re the ones that turn autonomy into reproducible, auditable safeguards. Evaluating a coding assistant without examining its sandbox model is taking on a security debt from day one. You’ve been warned.

G.

Meme of the day

Reply

or to participate.