The login you already have

The first thing co rem asks of a new user is to connect their mail. For someone who arrived from an ad, that is a second login for an account they had already connected that morning, in Codex or in Claude.

So we asked a narrower question: can co use the connection that is already there?

Codex answered first. Its app-server protocol, the one co already drives to run Codex as a coding harness, has a method called mcpServer/tool/call. It takes a session id, a server and a tool. A session can be started empty and marked ephemeral: no model turn, nothing written to the user's history. On the owner's Mac the account connectors came back as one server, codex_apps, with 537 tools. A Gmail search answered in 1.3 seconds with message ids and thread ids, which is everything a page needs to cite a mail.

Claude Code took longer to answer, and the answer was no. Its claude.ai connectors go through a proxy authorised by the user's own Claude login. The only way to call it without Claude would be to take that login's token and use it from another program, and that would be co pretending to be Claude Code with someone else's credentials. claude mcp serve exposes Claude Code's own tools, not the connectors. The supported path is a Claude turn, measured at 12 seconds and four cents for one search: fine for a question, not a way to read two years of mail.

So co mcp is three commands on top of Codex: ls, tools, call. Each tool already says whether it only reads. A read-only tool runs and prints its own data. Anything else (send, delete, create) prints what it would do and waits for --yes, because a connector acts as the user, and a command an agent can run should not send mail on a guess.

The lesson is in the order of the questions. We started from "connect your mailbox" as a step everyone takes. The better first question is "what can this person's tools already do". Here the answer was most of it.