So this appears to naively use DJB's NACL/crypto_box construction, which is a curious choice given the existence of OTR for messaging protocols which would handle things like session key negotiation and provide deniability.

First, If I'm reading the source correctly, they are doing public key encryption for every message. Which, ok, DJB was a fan of at least for DNSCurve, but is generally regarded somewhat dimly for efficiency reasons. So I guess this puts them on one extreme of the Bell Curve or the other. I wonder which?

[EDIT, removed point about nonce's in handshake]

Funnily enough, at first glance it looks like they covered at least some of the obvious issues: they do at least attempt to authenticate the session key and the crypto_box's use of a Nonce prevents replay and re-ordering attacks.

How do they handle video chat? Crypto_box won't work there naively sense packets will get lost and the nonce's won't be in sync.