Corpusly indexes a Google Drive for semantic search and serves it to Claude and other MCP clients, with no vendor server anywhere in the data path. A look at the local embedding pipeline built on a quantized, in-process transformers.js model, the two-transport MCP server it’s served over, and the engineering it took to make ’nothing leaves the machine’ actually true: a memory blowout fixed by sub-batching, a stdout-corruption constraint solved structurally, and an OAuth scope decision that doubled as a compliance strategy.
