Google Workspace holds the material most business work actually runs on. The thread where the decision was made, the spreadsheet the numbers came from, the folder the contract lives in, the calendar that says who is free on Thursday.
AI agents mostly cannot reach any of it. So people bridge the gap by hand: copying a thread into a chat window, pasting the model's answer back into a reply, downloading a sheet to ask a question about it, re-uploading the result. The agent is capable and the data exists, and the two never meet without a human carrying material between them.
This closes that gap. It exposes ten Workspace services as tools any MCP-speaking agent can call, so the agent reads the actual mailbox and writes the actual document rather than working from whatever was pasted in.
The design decision that matters is that it is built to be hosted once for an organisation rather than installed per laptop. That sounds like an infrastructure detail and is really a security one: the moment a server is shared, the question of whose mailbox an agent is reading stops being obvious, and has to be answered by the architecture instead of by trust.
