r1-eye: Ambient AI Eye for Rabbit R1

Vision system for the jailbroken Rabbit R1 (CipherOS + Magisk root). A PC drives everything via ADB over WiFi — no server on the device, no API keys on the R1.

Two modes: Roast Sentinel monitors coffee roasts with Artisan integration, and AI Eye handles general-purpose vision queries.

Roast Sentinel

Start the sentinel, then roast normally in Artisan. It captures photos synced to your roast events and analyzes bean color via Claude Vision.

run_r1-eye eye.py sentinel start --bean "Ethiopia Yirgacheffe"
run_r1-eye eye.py sentinel start --bean "Ethiopia Yirgacheffe" --stream   # with live video
run_r1-eye eye.py sentinel start --bean "Test" --debug                    # raw WebSocket logs

Workflow:

1. Start sentinel — WebSocket server starts on port 8765
2. Press ON in Artisan — connection established, camera opens
3. Press CHARGE — capture loop starts (T=0 reference); stream window opens if --stream
4. Captures fire on each event (CHARGE, DRY, FCs, FCe) plus timed intervals (30s/20s/10s by phase)
5. Press COOL END — sends DROP, session ends, log saved to captures/
6. Press OFF — sentinel reads the saved .alog and links roast_uuid and batch_nr into the session log

Session logs land in captures/sentinel_YYYY-MM-DD_HHMM.json and are picked up by the coffee-roasting analysis pipeline.

Artisan WebSocket configuration: see CLAUDE.md.

run_r1-eye eye.py sentinel status    # last session summary
run_r1-eye eye.py sentinel log       # last session full log

AI Eye

General-purpose vision queries — point the R1 camera at anything and ask about it.

run_r1-eye eye.py ask "What's on my desk?"
run_r1-eye eye.py look              # capture only (no AI query)
run_r1-eye eye.py status            # check device connectivity

Commands

Command	Description
eye.py ask <question>	Capture + Claude Vision query
eye.py look	Capture image only (camera test)
eye.py deploy	Push device scripts to R1
eye.py status	Check R1 device connectivity
eye.py sentinel start [--bean NAME] [--port PORT] [--stream] [--crack] [--debug]	Start roast sentinel
eye.py sentinel status	Last session summary
eye.py sentinel log	Last session full log
eye.py stream start [--resolution RES] [--bitrate RATE]	Standalone live stream (blocks until Enter)
eye.py stream stop	Kill running stream
eye.py stream status	Check if stream is running

All commands are run via run_r1-eye eye.py <command>.

Deployment

r1-eye is designed to run on a dedicated machine next to the roaster. Deploy via rsync (no git needed on the roaster):

./deploy.sh              # quick: .py files + device scripts (~1s)
./deploy.sh --full       # full: everything + reinstall deps

Configure SSH alias roaster in ~/.ssh/config for the deploy scripts.

Captures sync automatically at session end (via R1_PUSH_ADDRESS env var). Fallback:

./sync_captures.sh            # pull missing JSONs/images from roaster
./sync_captures.sh --dry-run  # preview only

Setup

Prerequisites

• Rabbit R1 with CipherOS + Magisk root, connected via WiFi ADB
• grab_power3 running on device
• adb installed on PC
• uv for Python package management
• Artisan 4.0 for roast sentinel integration
• scrcpy 3.x for live streaming (apt version too old — build from source, see CLAUDE.md)

Installation

# Clone and enter
git clone https://github.com/frogmoses/r1-eye.git
cd r1-eye

# Create venv and install deps
uv venv .venv --seed
uv pip install anthropic websockets numpy pillow

# Configure secrets
mkdir -p ~/.config/code-projects/r1-eye
cp .env.example ~/.config/code-projects/r1-eye/.env
chmod 700 ~/.config/code-projects/r1-eye
chmod 600 ~/.config/code-projects/r1-eye/.env
# Edit .env — fill in ANTHROPIC_API_KEY and R1_ADB_ADDRESS

# Create wrapper script (injects env vars from .env at runtime)
cat > ~/.local/bin/run_r1-eye << 'WRAPPER'
#!/usr/bin/env bash
set -euo pipefail
set -a
source ~/.config/code-projects/r1-eye/.env
set +a
exec "$PWD/.venv/bin/python" "$@"
WRAPPER
chmod +x ~/.local/bin/run_r1-eye

# Deploy capture scripts to R1
run_r1-eye eye.py deploy

Environment Variables

See .env.example for the full list. Required:

Variable	Purpose
ANTHROPIC_API_KEY	Claude Vision API key
R1_ADB_ADDRESS	R1 WiFi ADB address (e.g. 192.168.1.100:5555)
R1_PUSH_ADDRESS	Rsync target for session log push (e.g. user@devbox:~/r1-eye/captures/)
ARTISAN_SAVE_DIR	Artisan .alog save directory (default: ~/coffee-roasts)

Testing

For testing without the R1 or real Artisan:

.venv/bin/python fake_artisan.py           # simulated 10-min roast timeline
.venv/bin/python fake_artisan.py --fast    # compressed 30s timeline
run_r1-eye sim_sentinel.py                 # full simulation with reference images + real Vision API
run_r1-eye test_artisan_ws.py              # raw WebSocket debug server

Known Issues

• DROP button unavailable via WebSocket — Artisan's DROP action slot is used by Hottop safety commands. COOL END button sends DROP instead. Ctrl+C also works.
• scrcpy apt package too old — v1.25 doesn't support this device's Android version. Build v3.x from source (see CLAUDE.md).

For architecture, image pipeline details, device internals, and file-level documentation, see CLAUDE.md.