QVAC Can Run a Brain–Computer Interface. Here's What That Means.
QVAC can decode brain signals into text, entirely on-device. What BCI transcription is, and why local processing matters most here.
Buried in QVAC's capability list, between translation and vision-language-action, is a line that most people scroll straight past:
BCI: brain–computer interface transcription that decodes multi-channel neural signals into text.
On-device. No cloud. Via a customized Whisper backend.
Tether shipped brain–computer interface support in an open-source SDK, and almost nobody has written a word about it. Here's what it actually is and why it matters more than it first appears.
What BCI transcription actually does
A brain–computer interface reads electrical activity from the brain — typically via multiple electrode channels — and turns those signals into something a computer can use. BCI transcription specifically means decoding those neural signals into text.
The parallel to speech recognition is exact, which is why the implementation is interesting: QVAC decodes neural signals using a customized Whisper backend (bci-whispercpp). Whisper is the architecture that revolutionized speech-to-text. Neural signals, like audio, are multi-channel time-series data. Treating brain signal decoding as a transcription problem — the same shape as turning sound waves into words — is an elegant reframing.
This is not text-to-speech. It's not thought-reading in any science-fiction sense. It's the decoding of measured neural signals into text output, and it's a real, active field with real medical applications: restoring communication to people with ALS, locked-in syndrome, or severe paralysis.
Why on-device matters more here than anywhere else
Every argument for local AI gets sharper as data gets more personal. Your notes are personal. Your finances are more personal. Your health data is more personal still.
Neural data is the end of that line. There is no more intimate data than the electrical activity of your brain.
Consider what a cloud-based BCI would mean. Your neural signals — streamed, continuously, to a server. Stored. Logged. Subject to a privacy policy that can change, a breach that can happen, a subpoena that can arrive. A company would hold a record of raw signals from inside your skull.
With QVAC's approach, the decoding happens on the device. The signals never leave. There's no server to store them, because there is no server.
If you accept the argument for on-device AI anywhere, you have to accept it here. This is the case that makes the principle undeniable.
The latency argument
Privacy isn't the only reason. Tether's own vision document makes the point about physics rather than policy:
"An autonomous car does not ask the cloud for permission to brake; it reacts."
The same is true of a BCI. A neural interface that has to round-trip to a data center before producing output is a neural interface with network latency between intention and result. For assistive communication — someone with ALS composing a sentence — that lag is the difference between a usable tool and a frustrating one. For any real-time control application, it's disqualifying.
Local inference removes the round trip entirely. Thought to text, with no data center in between.
Where this fits in QVAC's larger picture
BCI doesn't appear in QVAC by accident. Look at what sits next to it in the capability list: VLA (vision-language-action) for robot control. Tether's stated vision talks about "10 billion humans, 10 billion autonomous machines, and a trillion AI agents" and describes QVAC as designed to run "from the most powerful industrial server to the smallest chip in a light bulb."
BCI and robotics are the two ends of embodied AI: reading signals from a nervous system, and sending signals to a machine body. Both demand the same three properties — low latency, high privacy, no cloud dependency — and both are impossible to do well on rented infrastructure in a data center.
Whether or not you find Tether's "decentralized mind" framing grandiose, the engineering follows a consistent logic: intelligence at the edge, because the edge is where the sensors and the actuators are.
Being honest about the state of it
Some appropriate calibration:
QVAC is a pre-1.0 SDK. The BCI capability is documented and shipped, but this is early-stage software in an early-stage field. Decoding neural signals into accurate text is a hard research problem, and no SDK makes it easy — the quality of any BCI system depends enormously on the electrodes, the signal quality, the training data, and the specific model.
QVAC provides the inference layer: a way to run neural-decoding models on-device, cross-platform, with the same interface you'd use for text or speech. It doesn't magically solve BCI. What it does is make it possible to run BCI decoding privately and locally, which — given the sensitivity of the data — may turn out to be the precondition for anyone deploying it responsibly at all.
Why we find this interesting
We're building private notes and private finance apps on QVAC (both coming soon), which is a long way from neuroscience. But the reason we chose QVAC is the same reason its BCI support exists: the conviction that the most personal data should be processed where it's created, not shipped to someone else's building.
BCI is that conviction taken to its logical end. If AI is going to read signals from inside your head, it had better be running in your hand.
Want the technical overview? See QVAC SDK Explained and the complete guide to QVAC's capabilities. BCI implementation lives in the bci-whispercpp package; verify current details in the official docs.