Buddy — the EEG AI Agent

Buddy is a 100% browser-based EEG copilot — and it is the AI agent. It pairs a live brain panel with a tool-using LLM that reads your neural state client-side, plus guided pattern training — all powered by Web Bluetooth / Web Serial. There is no server to run, no account to create, and no data leaves the tab.

Buddy is part of PiEEG Cloud. Where the dashboard (opens in a new tab) is for watching your signal, Buddy is for talking to it — asking questions in plain language, training custom mental-state detectors, and reading an honest, explained picture of your live brain state.

Getting started

The brain panel

Buddy reduces the raw stream into language-sized facts and renders them live:

The AI agent

The chat copilot is a tool-using LLM that runs a multi-turn loop over your live neural cascade — reading state, spectra, artifacts, connectivity, and your trained patterns through read-only tools — and answers in plain language.

It is 100% browser-based. You bring your own API key, and the agent talks to your chosen provider directly from the tab. There is no PiEEG server hop, and no raw EEG is ever sent to the model.

How it works

The agent runs a bounded tool-calling loop:

1. Your question is appended to the conversation.
2. The model streams a response; if it requests tools, the agent runs them against the live BuddyEngine.
3. Tool results are fed back and the model continues, up to a small iteration cap, until it answers.

The model pulls state through tools on demand — it is never streamed raw samples. This solves the rate mismatch between an LLM (~1 Hz of reasoning) and EEG (250–500 Hz of data).

Perception cascade (browser edition)

The agent reads a layered reduction of the signal — trimmed to the senses that exist in-browser (no server, no actuators, no notebook/session store):

Tools

The agent grounds every claim in a fresh tool call against the live engine:

Providers

Pick a provider in the Lobby or in Settings and paste an API key. Cloud providers stream over their standard APIs; local providers run on your own machine; Echo needs no key and is handy for a no-cost demo.

Example conversation

you > how's my signal?
buddy > All channels read "good", quality 0.97 — trustworthy.

you > am I focused or relaxed right now?
buddy > Relaxed, for you right now. Alpha is dominant and focus is low.
        (Readings are within-session, not clinical.)

you > which of my trained patterns is firing?
buddy > "meditation" is active (p≈0.88). It's a strong detector —
        leave-one-rep-out balanced accuracy 0.89, reading mostly O1/O2.

Pattern training

Buddy's most powerful sense is a trained pattern detector — a custom classifier for any mental state you can reproduce (meditation, focus, imagined movement, etc.).

Training is contrastive: you record a REST baseline and an ACTIVE state across several reps. Buddy then fits a per-pattern detector and scores it honestly:

• Leave-one-rep-out cross-validation (LORO-CV) splits by rep, not by sample, so temporal leakage cannot inflate the score.
• Balanced accuracy is reported so class imbalance cannot flatter the result.
• Channel importance is exposed via spatial (group-lasso) sparsity — you can see which electrodes the detector actually reads.

A trained detector is far more meaningful than the generic focus / relax indices, and the AI agent can list, detect, and explain your patterns on demand.

Honest metrics

Buddy is deliberate about what its numbers mean, and its system prompt keeps the agent candid:

• focus / relax / engagement are 0–1 values relative to this session's own range — described as "high/low for you right now", never absolute or clinical.
• A warming_up flag marks the settling period before enough variance is established; the agent says so when a reading is still settling.
• Pattern scores are leave-one-rep-out balanced accuracy — honest at the data scale you actually have.
• Poor signal quality or flagged channels are surfaced before conclusions are drawn.

Privacy

Everything stays in your browser tab:

• Raw EEG is decoded, filtered, and analyzed client-side — it is never uploaded to a PiEEG server.
• Trained patterns persist locally in your browser.
• Your API key and every agent request go directly from the tab to your chosen LLM provider — there is no PiEEG middleman. The model receives only the reduced, language-sized facts returned by tools, never raw EEG. You choose whether to persist the key.

Browser vs. server-side

Supported devices

Next steps