DJB: Network and Cryptography Library (efforts against timing attacks)
nacl.cr.yp.to
nacl.cr.yp.to
The fast and secure/non-exploitable (re-)implementation of core crypto algorithms is definitely useful. The restrained resource usage and no malloc dependency is something that I wish more libraries paid more attention to.
However, the stated target audience of the library seems odd. They say that the API is meant to abstract away the cryptographic details, and yet it still requires a crypto expert to go behind the API to set things up... which is fine, but if there is a crypto guy on a team, he could as easily cook up the abstraction interface around whatever crypto library they currently use. Unless I am missing something obvious, this abstraction layer and the whole focus on "usability" appears to be more of a hindrance that complicates access to the actually useful parts of the library (the high-speed crypto).
If you're not using the abstraction layer, the high-speed crypto bits you mention aren't useful anyway. The high speed crypto is djb's favourite primitives -- curve25519, salsa20, and poly1305 -- which is fine if you're only ever going to talk to other people using this library but makes it utterly useless otherwise since nobody else ever uses those primitives.