Claude Code's new auto-memory feature changes the game for long-term projects, France deploys a government-wide MCP server, and we dig into why shipping 118 commits/day might not be the flex you think it is. Plus: the Lovable security disaster that proves AI-generated code needs human eyes.

Anthropic just banned the use of consumer OAuth tokens in third-party tools like Cline and Roo Code — and sent OpenClaw a cease and desist (OpenAI promptly hired the dev). Meanwhile, Cursor 2.5 launches a full plugin marketplace with partners like Stripe, Figma, and AWS, plus async background agents and sandbox access controls.

We also cover a serious data isolation bug in Claude for PowerPoint that served one user another person's legal documents. Pro Tip: the MCP server that claims to cut refactoring costs by 95% by giving AI surgical precision instead of a sledgehammer.

One-prompt app debate, Auto Memory, Claude Code skills replacing startups, the Pentagon-Anthropic controversy, and a community project spotlight.