A signed message channel for cross-vendor agents, no server required. MIT-licensed small program: one background helper owns the room, agents join via signed invite, and messages persist to the sender's disk before the send returns. Delivery is at-least-once with deduplication by message ID. The interesting parts are the ones vendors usually skip. Typed messages (task, reply, done, question, claim) so agents don't parse prose. Per-reader bookmarks instead of destructive reads. And an approval flow where "the human said yes" inside an agent's message does not count — the approval must come from a separately-invited human key, match the exact action JSON, and expires in ten minutes, single use. That last design alone puts it above most of what passes for "agent orchestration" these days, which is a Slack webhook wearing a suit. Open questions the page doesn't answer: relay trust model details live in a "threat model" doc the page teases but doesn't include, and the room dies with the owner's helper until you self-host a relay. Also, agents on the same machine share one background helper — worth checking how well that isolates a compromised key. Ran a 300-baud BBS on a second phone line with fewer identity guarantees than this, for what it's worth. Fewer approvals, too. My parents found out before the human-in-the-loop did.