Two new security features land for Claude Managed Agents: self-hosted sandboxes keep agent execution inside your own infrastructure or a managed provider, while MCP tunnels let the agent reach services behind your security perimeter.
You can now swap tools, MCP servers, or vault IDs in a live session without restarting it. MCP tool outputs exceeding 100k tokens are automatically offloaded to Sandbox files.
New onboarding cookbooks cover Cloudflare, Daytona, Docker, Modal, and Vercel for developers setting up self-hosted sandboxes with Claude Managed Agents.
A new blog post covers the practical hard parts of running computer use reliably: getting click accuracy right, tuning thinking effort levels, managing context across long sessions, and recording UI demonstrations that Claude can replay later.
An Anthropic team member is on the ground at the Code with Claude London event.
The Claude Code team made a stop in London, likely for a meetup or event.
An Anthropic team member is running the workshop stage today and the breakout stage tomorrow at Code with Claude Extended in London. Worth tracking down if you're at the conference.
The terminology change is cosmetic on your end: spending limits, auto-reload settings, and any purchased credits all carry over untouched.
Andrej Karpathy has joined Anthropic, and bcherny is publicly excited about the new working relationship. No product details, just a warm welcome from an Anthropic team member.
A new team member named Andrej has joined the Claude Dev team.
