opencastor 2026.2.20.10

The Universal Runtime for Embodied AI

OpenCastor

The Universal Runtime for Embodied AI

56,006 lines of code · 2,233 tests · Python 3.10–3.12
Connect any AI model to any robot hardware through a single YAML config.

---

🚀 Install in 10 Seconds

curl -sL opencastor.com/install | bash

Other platforms

Windows 11 (PowerShell):

irm https://raw.githubusercontent.com/craigm26/OpenCastor/main/scripts/install.ps1 | iex

Docker:

docker compose up

Manual:

git clone https://github.com/craigm26/OpenCastor.git
cd OpenCastor
python3 -m venv venv && source venv/bin/activate
pip install -e ".[dev]"

Supports Linux, macOS (Apple Silicon & Intel), Windows 11, Raspberry Pi, Docker. Installer flags: --dry-run, --no-rpi, --skip-wizard

✨ What's New in v2026.2.20.10

• 8 AI providers — Anthropic, Google, OpenAI, HuggingFace, Ollama, llama.cpp, MLX, Claude OAuth proxy
• Tiered brain architecture — reactive rules → fast inference → deep planning
• Hailo-8 NPU vision — on-device YOLOv8 object detection at ~250ms
• OAK-D stereo depth camera — RGB + depth via DepthAI v3
• Community Hub — browse, install, and share robot recipes with castor hub
• Interactive wizard — provider selection, auth, messaging setup with recall of previous choices
• castor doctor — full system health diagnostic

🔄 Self-Improving Loop (Sisyphus Pattern)

Your robot learns from its mistakes. After each task, the Sisyphus Loop analyzes what happened, identifies failures, generates fixes, verifies them, and applies improvements automatically.

Episode → PM (Analyze) → Dev (Patch) → QA (Verify) → Apply

• Disabled by default — opt-in via castor wizard or YAML config
• 4 cost tiers — from $0 (local Ollama) to ~$5-15/mo (Claude)
• Auto-apply preferences — config tuning only, behavior rules, or manual review
• Rollback — undo any improvement with castor improve --rollback <id>
• ALMA consolidation — cross-episode pattern analysis for deeper learning

castor improve --episodes 10    # Analyze last 10 episodes
castor improve --status         # View improvement history

🧠 Tiered Brain Architecture

OpenCastor doesn't send every decision to a $0.015/request cloud API. Instead, it routes through three layers — only escalating when needed:

┌─────────────────────────────────────────────────────────┐
│                    MESSAGING LAYER                       │
│         WhatsApp · Telegram · Discord · Slack            │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│                   API GATEWAY                            │
│            FastAPI · REST · Webhooks · JWT               │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│               TIERED BRAIN STACK                         │
│                                                          │
│  ┌────────────────────────────────────────────────────┐  │
│  │  Layer 3: PLANNER         Claude Opus · ~12s       │  │
│  │  Complex reasoning, multi-step plans, novel tasks  │  │
│  ├────────────────────────────────────────────────────┤  │
│  │  Layer 2: FAST BRAIN      HF / Gemini · ~500ms    │  │
│  │  Classification, simple Q&A, routine decisions     │  │
│  ├────────────────────────────────────────────────────┤  │
│  │  Layer 1: REACTIVE        Rule engine · <1ms       │  │
│  │  Obstacle stop, boundary enforce, emergency halt   │  │
│  └────────────────────────────────────────────────────┘  │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│                  PERCEPTION                              │
│  Hailo-8 NPU (YOLOv8) · OAK-D Depth · Camera · IMU    │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│               RCAN SAFETY KERNEL                         │
│    Physical bounds · Anti-subversion · Audit chain       │
└──────────────────────┬──────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────┐
│                  DRIVER LAYER                            │
│       PCA9685 · Dynamixel · GPIO · Serial · I2C         │
└──────────────────────┬──────────────────────────────────┘
                       │
                  [ YOUR ROBOT ]

Cost-effective by default. The reactive layer handles 80% of decisions at zero cost. The fast brain handles another 15%. The planner only fires for genuinely complex tasks.

🤖 Supported AI Providers

Provider	Models	Latency	Best For
Anthropic	claude-opus-4-6, claude-sonnet-4-5	~12s	Complex planning, safety-critical reasoning
Google	gemini-2.5-flash, gemini-2.5-pro, gemini-3-flash-preview	~500ms	Multimodal, video, speed
OpenAI	gpt-4.1, gpt-4.1-mini, gpt-5	~2s	Instruction following, 1M context
HuggingFace	Transformers, any hosted model	~500ms	Fast brain layer, classification
Ollama	llava:13b, any local model	Varies	Privacy, offline, zero cost
llama.cpp	GGUF models	~200ms	Edge inference, Raspberry Pi
MLX	Apple Silicon native (mlx-lm, vLLM-MLX)	~50ms	Mac M1-M4, 400+ tok/s
Claude OAuth	Proxy-authenticated Claude	~12s	Team/org deployments

Swap providers with one YAML change:

agent:
  provider: "anthropic"
  model: "claude-opus-4-6"

👁️ Vision & Perception

Hailo-8 NPU — On-Device Object Detection

No cloud API calls needed. The Hailo-8 neural processing unit runs YOLOv8 locally:

• 80 COCO object classes — people, vehicles, animals, furniture, and more
• ~250ms inference — fast enough for real-time obstacle avoidance
• Zero API cost — all processing happens on the edge

OAK-D Stereo Depth Camera

RGB + depth streaming via DepthAI v3. Get 3D spatial awareness for navigation, manipulation, and mapping.

perception:
  camera:
    type: "oakd"
    depth: true
    resolution: "1080p"
  npu:
    type: "hailo8"
    model: "yolov8n"
    confidence: 0.5

🛡️ Safety First

OpenCastor implements defense-in-depth safety, inspired by ContinuonAI principles and fully RCAN spec compliant:

Layer	What It Does
Physical Bounds	Workspace limits, joint constraints, force capping
Anti-Subversion	Prompt injection defense, input sanitization
Work Authorization	Dangerous commands require explicit approval
Tamper-Evident Audit	Hash-chained logs at /proc/safety — any tampering is detectable
Emergency Stop	Hardware and software e-stop, reactive layer < 1ms

castor audit --verify        # Verify audit chain integrity
castor approvals             # View/approve dangerous commands
castor privacy --config r.yaml  # Show sensor access policy

📦 Quick Start

1. Install & Configure

curl -sL opencastor.com/install | bash
castor wizard

The wizard walks you through provider selection, API keys, hardware config, and optional messaging setup (WhatsApp/Telegram). It remembers your previous choices.

2. Run

castor run --config my_robot.rcan.yaml

3. Open the Dashboard

http://localhost:8501

4. Diagnose Issues

castor doctor

Minimal Python Example

from castor.providers import get_provider
from castor.drivers.pca9685 import PCA9685Driver

brain = get_provider({"provider": "anthropic", "model": "claude-opus-4-6"})
driver = PCA9685Driver(config["drivers"][0])

while True:
    frame = camera.capture()
    thought = brain.think(frame, "Navigate to the kitchen, avoid obstacles.")
    if thought.action:
        driver.move(thought.action["linear"], thought.action["angular"])

🏪 Community Hub

Browse, install, and share robot recipes:

castor hub search "delivery bot"
castor hub install @alice/warehouse-picker
castor hub publish my_robot.rcan.yaml

Recipes are shareable RCAN configs — complete robot personalities with perception, planning, and safety settings bundled together.

🔧 CLI Reference

Setup & Config

castor wizard                          # Interactive setup wizard
castor quickstart                      # Wizard + demo in one command
castor configure --config robot.yaml   # Interactive config editor
castor install-service --config r.yaml # Generate systemd unit file
castor learn                           # Step-by-step tutorial
castor doctor                          # Full system health check
castor fix                             # Auto-fix common issues

Run & Monitor

castor run --config robot.yaml         # Perception-action loop
castor run --config robot.yaml --simulate  # No hardware needed
castor gateway --config robot.yaml     # API gateway + messaging
castor dashboard                       # Streamlit web UI
castor demo                            # Simulated demo
castor shell --config robot.yaml       # Interactive command shell
castor repl --config robot.yaml        # Python REPL with robot objects
castor status                          # Provider/channel readiness
castor logs -f                         # Structured colored logs
castor benchmark --config robot.yaml   # Performance profiling

Hardware & Recording

castor test-hardware --config robot.yaml  # Test motors individually
castor calibrate --config robot.yaml      # Interactive calibration
castor record --config robot.yaml         # Record a session
castor replay session.jsonl               # Replay a recorded session
castor watch --gateway http://127.0.0.1:8000  # Live telemetry

Hub & Fleet

castor hub search "patrol bot"         # Browse community recipes
castor hub install @user/recipe        # Install a recipe
castor hub publish config.yaml         # Share your recipe
castor discover                        # Find RCAN peers on LAN
castor fleet                           # Multi-robot status (mDNS)

Safety & Admin

castor approvals                       # View/approve dangerous commands
castor audit --since 24h               # View audit log
castor audit --verify                  # Verify chain integrity
castor privacy --config robot.yaml     # Sensor access policy
castor token --role operator           # Issue JWT
castor upgrade                         # Self-update + health check

🏗️ Supported Hardware

Pre-made RCAN presets for popular kits, or generate your own:

Kit	Price	Preset
Waveshare AlphaBot / JetBot	~$45	presets/waveshare_alpha.rcan.yaml
Adeept RaspTank / DarkPaw	~$55	presets/adeept_generic.rcan.yaml
SunFounder PiCar-X	~$60	presets/sunfounder_picar.rcan.yaml
Robotis Dynamixel (X-Series)	Varies	presets/dynamixel_arm.rcan.yaml
Hailo-8 + OAK-D Vision Stack	~$150	presets/hailo_oakd_vision.rcan.yaml
DIY (ESP32, Arduino, custom)	Any	Generate with castor wizard

🏫 STEM & Second-Hand Hardware

OpenCastor explicitly supports the parts that students, educators, and hobbyists actually have — donated kits, school surplus, eBay finds, and sub-$50 Amazon staples. If you found it at Goodwill, a school auction, or a makerspace scrap bin, there's probably a preset for it.

Kit	Typical New Price	Where to Find Used	Preset
LEGO Mindstorms EV3	~$300 new	School surplus, eBay $30-80	presets/lego_mindstorms_ev3.rcan.yaml
LEGO SPIKE Prime	~$320 new	STEM program donations, eBay $80-150	presets/lego_spike_prime.rcan.yaml
VEX IQ System	~$250 new	Robotics team surplus, school auctions $50-120	presets/vex_iq.rcan.yaml
Makeblock mBot	~$50 new	eBay $10-25, Amazon Warehouse	presets/makeblock_mbot.rcan.yaml
Arduino + L298N (DIY)	~$8-15 total	Makerspace bins, AliExpress	presets/arduino_l298n.rcan.yaml
ESP32 + Motor Driver (DIY)	~$6-12 total	AliExpress, hackerspaces	presets/esp32_generic.rcan.yaml
Yahboom ROSMASTER X3	~$150-200	Amazon Warehouse, eBay	presets/yahboom_rosmaster.rcan.yaml
Elegoo Tumbller / Smart Car	~$35-40 new	Amazon Warehouse $15-25, eBay	presets/elegoo_tumbller.rcan.yaml
Freenove 4WD Car (Pi-based)	~$40 new	eBay $15-25 (Pi not included)	presets/freenove_4wd.rcan.yaml
Cytron Maker Pi RP2040	~$10 new	Hackerspaces, STEM lab surplus	presets/cytron_maker_pi.rcan.yaml

🔍 Not sure what you have? See the Hardware Identification Guide for a decision tree: "I found this at a thrift store, now what?"

Tips for Second-Hand Hardware

• Test first, code later. Run castor test-hardware --config <preset>.rcan.yaml to verify each motor and sensor before writing any autonomy code.
• Cables are the most common failure point. LEGO connector cables, USB-B ports, and servo leads are all cheap to replace.
• Clone boards are fine. Arduino Uno clones with CH340 USB chips work perfectly. You may need to install the CH341SER driver on Windows.
• Battery health matters. Test battery packs under load — many donated robots have degraded cells that drop voltage and confuse motor drivers.
• Community firmware exists for almost every kit. Check the firmware/ directory in this repo for Arduino sketches and MicroPython scripts.

🤝 Contributing

OpenCastor is fully open source (Apache 2.0) and community-driven.

• Discord: discord.gg/jMjA8B26Bq
• Issues: GitHub Issues
• PRs: See CONTRIBUTING.md
• Twitter/X: @opencastor

Areas we need help with: driver adapters (ODrive, VESC, ROS2), new AI providers (Mistral, Grok, Cohere), messaging channels (Matrix, Signal), sim-to-real (Gazebo / MuJoCo), and tests.

📄 License

Apache 2.0 — built for the community, ready for the enterprise.

---

Built on the RCAN Spec by Craig Merry