Connecting AI assistants to a real mailbox: what they can read, what they can send, and where the limits actually sit.
Letting an AI assistant into a mailbox is a different problem from letting it read a web page. Mail is live, it belongs to somebody, and a mistake in it reaches a real person. The interesting questions are rarely about the model: they are about what the connector lets it name, what it refuses, and whether you can tell afterwards that a machine did the work.
These articles cover that ground from the build side. Press Forge and the Mailbox MCP server are both BSolve IT Limited, so the examples here come from a tool surface real customers use on live mail, including the parts we got wrong and corrected.
Ask an AI to email you a PDF and it either says it can't, or it starts typing the file out as base64 and runs out of reply. That is not the model being bad at files. MCP tool arguments are JSON, and JSON has no file type, so every server hits the same wall separately. Here is the gap in the spec, what base64 really costs (including the figure we published wrong for a month), and the four routes we built instead.
Press Forge
