EPOC X → keystroke-aligned EEG decoder → Hermes Agent → Unreal Engine MCP
Brain2Qwerty-inspired BCI stack for 14-channel EMOTIV EPOC X — not compatible with Meta lab EEG checkpoints.
NeuroGlyph Forge is an open research-to-demo pipeline for mind-linked typing and intent control:
- Record raw EEG (Cortex Premium) + keyboard timestamps
- Train
TinyB2Q/EEGNeton aligned 500 ms epochs (14×25 @ 50 Hz) - Decode live with confidence gates
- Reason in Hermes Agent (MCP tools)
- Act in Unreal via editor MCP (
MoveLeft,SpawnDebugActor, …)
Part of the broader MindBot / synergetic cognition story — this repo is the sensory-motor slice (see docs/MINDBOT_INTEGRATION.md).
Hand / zone / intent from EEG → Hermes → Unreal scene action.
Progressive tasks: hand → zone → char29 / intent.
git clone https://github.com/TheMindExpansionNetwork/neuroglyph-forge.git
cd neuroglyph-forge
python -m venv .venv
source .venv/Scripts/activate # Windows Git Bash
pip install -e .
pytest tests/ -q
bash scripts/pipeline.sh hand
python scripts/live_bci_demo.py --seconds 3Live headset: docs/CORTEX_SETUP.md · Unreal: docs/UNREAL_SETUP.md · Architecture: docs/SYSTEM.md · Brain map: docs/BRAINMAP.md · interactive HTML
Train: docs/TRAINABLE_DATASETS.md · Modal: docs/MODAL_TRAINING.md · Adapt: docs/ADAPTIVE_ORCHESTRATION.md · Why / English: docs/WHAT_THIS_DOES_FOR_YOU.md · First session: docs/FIRST_SESSION_PROTOCOL.md · 4070: docs/LOCAL_GPU.md · HF upload: docs/HF_UPLOAD.md · data/catalog.json
| Asset | Repo |
|---|---|
| Fine-tuned hand model | NeuroGlyph-EPOC-Typing-v1 |
| Raw session staging (EN) | neuroglyph-epoc-typing-en-v1 |
Weights and session CSVs live on HF — not committed to GitHub. Clone code here, pull checkpoints from HF or train locally (scripts/run_swarm_pipeline.py).
| Path | Role |
|---|---|
neuroglyph_recorder/ |
Cortex JSON-RPC client, keylogger, session CSV |
neuroglyph_data/ |
Epochs, EMOTIV CSV import, synthetic CI data |
neuroglyph_models/ |
TinyB2Q, EEGNet, task heads |
neuroglyph_train/ |
Train, eval, finetune, TorchScript export |
neuroglyph_agent/ |
Hermes MCP tools + live decoder |
neuroglyph_unreal/ |
Prediction → UE blueprint events |
scripts/ |
pipeline.sh, live_bci_demo.py, collect_session.py |
In-repo stdio server (train, preprocess, Unreal bridge):
# ~/.hermes/config.yaml — snippet in configs/hermes-mcp-snippet.yaml
mcp_servers:
neuroglyph:
command: "D:/neuroglyph-forge/.venv/Scripts/python.exe"
args: ["-m", "neuroglyph_agent.hermes_stdio_mcp"]Comfy Cloud (promo art, generative workflows): docs/COMFY_MCP.md — Comfy Cloud MCP.
hermes mcp test comfy-cloud # expects X-API-Key header| Source | Use |
|---|---|
| SpanishBCBL | Pipeline / CER reference only |
| NeuroGlyph-EPOC-Typing-v1 (you record) | Required for real accuracy |
Regenerate with OpenAI gpt-image-2 (fallback gpt-image-1.5 / gpt-image-1):
python scripts/generate_brand_assets.py --all
python scripts/generate_brand_assets.py --brainmapOutputs under assets/brand/. If OpenAI billing limit hits, use Hermes mcp_comfy_cloud_partner_generate with prompts in assets/brand/manifest.json.
Independent implementation. Brain2Qwerty is CC BY-NC 4.0; do not redistribute their weights with this repo.
- Inspired by brain2qwerty
- Agent layer: Hermes Agent
- Visual promo: OpenAI Images API (
gpt-image-2) + Comfy Cloud MCP