Inky Ganbold’s Brain-to-AI Tool Publishes Its Failures Alongside Its Results

Inky Ganbold
Written By:
Arundhati Kumar
Published on: 
Updated on: 

The open-source bci-mcp project connects live EEG signals to AI assistants through the Model Context Protocol. Inky Ganbold tested it on a public research dataset, the PhysioNet EEG Motor Movement/Imagery dataset, covering 109 people, and published both the results that held up and the detector that did not.

One result from Inky (Enkhbold) Ganbold’s bci-mcp benchmark looked encouraging: when participants closed their eyes, the tool’s ‘calm’ reading increased in 105 of 109 EEG recordings from the public PhysioNet EEG Motor Movement/Imagery dataset. Another result was considerably less flattering. The project’s blink detector misfired on roughly 40 percent of eyes-closed windows. Ganbold published both, making the limitation part of the project record instead of presenting only the strongest result.

That contrast is central to the way Ganbold approaches neurotechnology. A hardware and embedded-systems engineer, he created bci-mcp to stream live electroencephalography data into AI assistants through the Model Context Protocol, or MCP. The tool gives an assistant such as Claude access to live brain-signal data as an external input, but Ganbold has been careful to keep his claims narrower than the broader language often associated with brain-computer interfaces.

“I think the responsible thing is to show where a system works and where it breaks,” Ganbold said. “If a detector fails often enough to matter, that should be visible to the person using it instead of hidden behind the strongest demo.”

EEG makes that distinction especially important because the signals themselves are difficult to interpret. Electrical activity measured at the scalp is small, susceptible to noise, and influenced by movement and other artifacts. A measurable change does not automatically support a larger claim about what a person is thinking. Ganbold’s benchmark therefore focused on specific behaviors that could be checked across a defined dataset instead of relying on a single demonstration.

For the ‘calm’ measure, he compared the system’s response during eyes-open and eyes-closed conditions across recordings from 109 people in the public PhysioNet EEG Motor Movement/Imagery dataset. The reading increased with eyes closed in 105 cases, giving him a concrete performance result across most of the dataset. The blink detector produced a different picture, with false detections appearing in roughly 40 percent of eyes-closed windows. Ganbold reported that result as a limitation instead of treating the detector as finished.

“The useful question is not whether a graph moves,” Ganbold said. “The useful question is whether it moves consistently for the reason you think it does, and whether you have tested enough examples to know where that interpretation stops being reliable.”

bci-mcp is at the center of two developing technologies. Consumer EEG devices are making non-invasive brain-signal collection more accessible, while MCP gives AI assistants a standardized way to connect with external tools and data. Ganbold’s project links those layers by allowing EEG information to become an input an AI assistant can receive and process.

That connection changes the role of the AI assistant. Instead of responding only to text supplied in a conversation, the assistant can receive changing information from an external source. In bci-mcp, that source is EEG data. The value of the connection therefore depends not only on whether the data can be transmitted, but on whether the measurements being exposed to the assistant have been characterized honestly enough for developers to understand what they represent.

The technical connection is only part of the story. The benchmark also illustrates a recurring problem in neurotechnology: a system can produce an interesting output without proving every interpretation attached to it. Ganbold’s decision to publish both the strong and weak results gives developers a clearer boundary between demonstrated behavior and unfinished capability. It also reduces the risk that a weak detector will be treated as reliable simply because it appears alongside a stronger feature in the same project.

The project is publicly available through npm and PyPI and has accumulated about 1,000 downloads. It has also been listed in more than 20 MCP directories. Those figures show that developers are finding and downloading  the tool, but they do not establish whether each detector performs reliably. Ganbold treats the benchmark as a separate question because adoption and technical validity measure different things.

That distinction matters in a field where demonstrations can easily be described with language that outruns the evidence. Ganbold has been explicit that bci-mcp does not establish an ability to read minds. Instead, it streams EEG data into an AI environment and provides measurable outputs that can be tested against known conditions. The benchmark gives developers a clearer picture of which parts of the system have behaved consistently and which still require more work.

“I would rather publish a limitation that makes the project look less impressive than let somebody build on a capability that is not actually there,” Ganbold said. “If the evidence improves later, then the claim can improve with it.”

The 109-person dataset gives that approach substance. A result seen in 105 participants is stronger than a single successful demonstration, while a detector that misfires across roughly 40 percent of a defined condition clearly signals unfinished work. For a developer evaluating whether to use or extend the project, those two numbers serve different purposes: one supports confidence in a measured behavior, while the other marks an area where additional testing or a different approach may be necessary.

Ganbold’s broader technical background helps explain why he approaches the project this way. His work spans embedded hardware and software, so he is accustomed to separating measured performance from expectation. In bci-mcp, that discipline appears in the way he reports results, with each capability standing on the evidence collected for that capability instead of inheriting credibility from the project as a whole.

The project therefore contributes more than a connection between EEG and AI assistants. It offers a practical example of how neurotechnology can be presented without turning every successful signal into a larger claim. The most useful result may be the contrast itself, because it shows which parts of the system behaved consistently and which still require improvement.

As consumer neurotechnology and AI tooling continue to converge, transparent benchmarking will matter as much as technical novelty. Ganbold’s approach with bci-mcp is restrained: define the behavior, test it across a meaningful dataset, report the numbers, and leave unsupported conclusions out. In a field where credibility depends heavily on what the data can actually show, the failures belong in the story too.