Exploring Semantic MCP
I built an MCP server with zero API keys, zero databases, and zero state. A static file and five functions. It works better than it should.
Essays and explorations on AI, Internet Services, and Customer Experience Design.
I built an MCP server with zero API keys, zero databases, and zero state. A static file and five functions. It works better than it should.
AI tooling is making software teams faster than ever - but the processes they're running were designed for when speed was the problem. What happens when building is cheap and the bottleneck moves somewhere else?
PSA for anyone dealing with SaaS tools that have good reporting UIs but garbage APIs: Claude Chrome + Claude Code might be your workaround.
Claude Code skills and spawned subagents solve the same problem—selective context loading, but for different architectures. One works inside a session, the other across processes.
Today’s models fail loudly and visibly. That makes their limits easy to study and their mistakes easy to catch. Future systems will not be so forgiving.
I tried to see if I could get better results from combining the effort of three AI assistants, but the results weren't all that satisfying. But I learned a few things.
A personal site feels right again, so I rebuilt one with a simple GitHub and Netlify workflow and a bit of nostalgia for how the web used to work. If you’re curious how it came together or why I bothered, read on...