Last night an agent put 16 quiz photos onto a Lectern page through WebMCP. It worked. It was the wrong architecture.
WebMCP tools are still strings. So the pixels rode JSON: +33% base64, ~64 CDP round-trips. The model should never see the file.
We’ll probably extract one more library from Lectern for that: AMDP. Cite a hash, bind it to the page, move bytes on a side channel. Chunked JSON stays the fallback, not the protocol.
Log in or sign up for Devpost to join the conversation.