The Open-Source BCI Stack

It helps to read the ecosystem as a stack. Hardware designs at the base; acquisition libraries that read samples from devices; streaming protocols that keep clocks aligned; processing packages that clean and transform signals; decoding libraries and benchmarks; application frameworks that close the loop; and, alongside all of them, the data standards and archives that make results reusable.

Signal processing for non-invasive data is the deepest layer. MNE-Python has been developed in the open since 2011 and published a release last month. EEGLAB and FieldTrip serve the MATLAB community. Lab Streaming Layer has become the default way to synchronise devices.

For EEG decoding, MOABB gives the field a shared way to evaluate methods across datasets, and libraries such as pyRiemann and Braindecode provide strong, comparable baselines.

Closed-loop, real-time decoding for implanted systems is the least developed layer in the open. BRAND, built by BrainGate collaborators, and ezmsg are capable and young. Clinical-grade real-time infrastructure is mostly proprietary.

Tooling for training models across many sessions and subjects is newer still, and depends on standardised access to data that only a few groups hold.

Almost none of this software is a commercial product. It is sustained by research grants, institutional support and volunteers, while companies and laboratories around the world depend on it.

That is a structural fragility and an opportunity. Modest, sustained investment in the open stack would benefit every team building on it.