Online BCI experiments ask a lot of their software at once: pull in several physiological signals live, train a decoder, run it for real-time inference, and keep enough records that someone could reproduce the whole thing later. Dendrite is a Python application that puts those jobs in one place while staying open to modification.

It records multiple streams, each at its own sampling rate, and runs several processing modes against them concurrently. A decoder can be loaded pre-trained or fitted mid-session without stopping the pipeline, and the same recordings later drive offline training. Everything is logged to a database, so any deployed decoder traces back to its configuration and source recordings. The experiment itself stays outside, talking to Dendrite over the network in whatever language you like. The authors validate it end to end on in-house and public datasets.

It's released under GPL-3.0, and the paper covers the architecture the abstract only sketches.