Product updates, build decisions, field notes. One post per week. Short when the week is loud, long when the week is quiet.
An MCP server for a real app is mostly not about MCP. The hard parts: one ops layer across three transports, eventing bugs only real clients trigger, and why our 10 ms of overhead is noise next to the network.
Three decisions made the marketing site fast: Specific.dev for deploys, Astro 6 for the user experience, typed markdown for content. Here is what each buys us, and where the stack still bites.
Slack-style chat puts the server in the middle of every interaction. We picked Zero, a local-first sync engine, and worked the architecture inwards from the user. Here is what changed, and why.
