During the budget announcement at the end of the October 2026, Chancellor John Healey did what almost all of us do nowadays; he talked about AI. He said a lot of things that gave me pause to think, but one thing stuck out...
"When the NHS uses AI to treat patients, we want that data here, when our armed forces use AI to keep us safe, we want that data here."
Almost a year ago, while working on an open source project at Mozilla AI named mcpd (also a plugin system and proxy), I became frustrated in my dealings with the Model Context Protocol (MCP). So, like a good keyboard warrior, I wrote about it on the internet.
Since that time, a lot has changed in the fast-moving world of AI and with MCP. A few times over these months I thought: I should do a follow-up post, but then work got in the way. But fear not... the time has come.
Every AI model has blind spots. It might overlook context, lean toward certain patterns, or fill gaps with confident guesses. When you're using an AI coding agent to help with architecture decisions or code review, you're getting one perspective from one model. That's fine for straightforward work. But for decisions that shape the long-term direction of a codebase, one perspective isn't enough.
The Star Chamber is a skill for Claude Code that fans out code reviews and design questions to multiple LLM providers simultaneously, aggregates their feedback, and presents consensus-based recommendations. Think of it as a senior engineering council that reviews the same code independently, then you get a summary of where they agree and where they don't.
Last week I wrote about building toward sovereign AI - infrastructure that keeps your data local while letting AI actually help you. Today the repositories are public. Here's what's in them...
I've been spending evenings and weekends on something that might sound idealistic: infrastructure for personal data sovereignty. But I'm not building it because I think it'll definitely work. I'm building it because the alternative - doing nothing - seems worse.
I've written a follow-up to this post. It corrects some things I got wrong here and covers what has changed in MCP since.
The Model Context Protocol (MCP) promises to be the "USB-C" of LLM integrations; simple, universal, and "extremely easy to build." Yet a fundamental disconnect exists between its security guidance and how authentication works in production systems. This isn't just a theoretical concern; it's creating an ecosystem where even official implementations can't follow the stated best practices.