diff --git a/docs/conferencing.md b/docs/conferencing.md
index e69de29bb..874ce4cfd 100644
--- a/docs/conferencing.md
+++ b/docs/conferencing.md
@@ -0,0 +1,52 @@
+# VoIP Conferencing
+
+This is a draft proposal for a naive voice/video conferencing implementation for
+Matrix clients.  There are many possible conferencing architectures possible for
+Matrix (Multipoint Conferencing Unit (MCU); Stream Forwarding Unit (SFU); Peer-
+to-Peer mesh (P2P), etc; events shared in the group room; events shared 1:1;
+possibly even out-of-band signalling).
+
+This is a starting point for a naive MCU implementation which could provide one
+possible Matrix-wide solution in  future, which retains backwards compatibility
+with standard 1:1 calling.
+
+ * A client chooses to initiate a conference for a given room by starting a
+   voice or video call with a 'conference focus' user.  This is a virtual user
+   (typically Application Service) which implements a conferencing bridge.  It
+   isn't defined how the client discovers or selects this user.
+
+ * The conference focus user MUST join the room in which the client has
+   initiated the conference - this may require the client to invite the
+   conference focus user to the room, depending on the room's `join_rules`. The
+   conference focus user needs to be in the room to let the bridge eject users
+   from the conference who have left the room in which it was initiated, and aid
+   discovery of the conference by other users in the room.  The bridge
+   identifies the room to join based on the user ID by which it was invited.
+   The format of this identifier is implementation dependent for now.
+
+ * If a client leaves the group chat room, they MUST be ejected from the
+   conference. If a client leaves the 1:1 room with the conference focus user,
+   they SHOULD be ejected from the conference.
+
+ * For now, rooms can contain multiple conference focus users - it's left to
+   user or client implementation to select which to converge on.  In future this
+   could be mediated using a state event (e.g. `im.vector.call.mcu`), but we
+   can't do that right now as by default normal users can't set arbitrary state
+   events on a room.
+
+ * To participate in the conference, other clients initiates a standard 1:1
+   voice or video call to the conference focus user.
+
+ * For best UX, clients SHOULD show the ongoing voice/video call in the UI
+   context of the group room rather than 1:1 with the focus user.  If a client
+   recognises a conference user present in the room, it MAY chose to highlight
+   this in the UI (e.g. with a "conference ongoing" notification, to aid
+   discovery).  Clients MAY hide the 1:1 room with the focus user (although in
+   future this room could be used for floor control or other direct
+   communication with the conference focus)
+
+ * When all users have left the conference, the 'conference focus' user SHOULD
+   leave the room.
+
+ * If a conference focus user joins a room but does not receive a 1:1 voice or
+   video call, it SHOULD time out after a period of time and leave the room.