opencastor.com

What It Is

OpenCastor positions itself as a “Physical AI universal runtime.” Its goal is to connect robot hardware, AI models, perception, safety, and communication through a single RCAN configuration file. It is not a typical chatbot app, but open-source infrastructure for real robots, mobile platforms, sensors, and actuators. It uses the Apache 2.0 license and provides GitHub, PyPI, documentation, and CLI-based installation options.

Core Capabilities

At its core is a three-layer cognitive architecture: the Reactive Layer handles hardware-level safety reflexes such as emergency stop and obstacle avoidance, runs in under 1ms, and does not depend on AI; Fast Brain uses models from Hugging Face, Ollama, llama.cpp, and similar sources for real-time perception and quick decisions, at around 100ms; Planner calls Claude, GPT-4.1, and Gemini for long-horizon reasoning and task decomposition, at around 2 seconds. On the hardware side, it supports USB cameras, depth cameras, LiDAR, IMU, motor controllers, OAK-D, Intel RealSense, Hailo-8, and HLabs ACB v2.0. Semantic perception uses local CLIP embeddings by default, with an option to upgrade to Gemini Embedding 2.

Pricing and Integrations

OpenCastor itself is described as “Free forever.” The Primary Brain can run for free via Apple Foundation Models, Hugging Face’s free inference API, Ollama, and llama.cpp; for more complex planning, users can bring in paid third-party models from Anthropic, Google, OpenAI, and others as needed. Claude OAuth can reuse an existing Claude Pro subscription. The CLI supports installation, hardware scanning, interactive configuration, and robot operation. It also implements the open RCAN protocol, Robot Registry, and ROBOT.md, making it suitable for teams that need auditable and registrable robot identities.

Pros, Cons, and Who It’s For

The main advantages are that it is open source, avoids model lock-in, runs across Raspberry Pi, Jetson, x86 Linux, and macOS, and places safety, auditing, and Human-in-the-Loop workflows directly into the architecture. Community recipes cover scenarios such as home patrols, farm inspections, and classroom Q&A. The limitation is that it remains a robotics development tool: users need to understand hardware, sensors, models, and configuration. Output quality depends heavily on the chosen model, hardware, and use case. Chinese UI, Chinese documentation, and Chinese-language capabilities are not clearly specified.

Access from China

The scraped text does not specify network accessibility or payment options for mainland China. Since it may rely on services such as GitHub, Hugging Face, Claude, Gemini, and OpenAI, usage in China is likely to be constrained by network access and third-party account requirements. A local Ollama/llama.cpp setup would be more controllable. Comparable alternatives include ROS 2, NVIDIA Isaac ROS, or self-built robotics stacks using local models.