Three notes from January 2026, written while I was building the first composable MCP servers over a data platform: warehouse access through Trino, object storage access through S3, and the conclusion I reached about the protocol after shipping both. The short version of that conclusion is that MCP is flawed and worth building on anyway, because context was always the hard problem and the protocol at least forces you to solve it in the open.

These are the starting point for everything that followed. The protocol itself is pulled apart in MCP Without the SDK, and the engineering around it is in MCP by Design.