MD-OS aims to develop a persistent agentic operating layer for advanced automation.
Linux gave users freedom and control over the machine. MD-OS gives agents freedom and control over their operational context: files, rules, memory, policies, execution, verification, and logs. Free agents need a free operating system.
The goal is to coordinate software, devices, robotic systems, memory, policies, and feedback loops through a structured operational context that is readable, auditable, correctable, and reusable across sessions.
Instead of treating automation as isolated commands or one-shot prompts, the project explores automation as an evolving operational system: one that can preserve context, select bounded actions, observe outcomes, learn from feedback, and safely orchestrate heterogeneous environments.
This public lab documents selected experiments at a high level. Source code, connector configurations, private runtime state, and internal implementation details remain confidential.
Latest experiment
Diary Day 1 - Self-Observation With a New Robotic Arm
Speedy Evolv begins learning how to use a new mechanical arm attachment by observing its own movements through a webcam.
This experiment combines hardware autodiscovery, action, perception, feedback, and adaptation. The system detects and models a new physical extension, performs controlled arm movements, visually observes the real-world outcome, and begins building the foundation for self-correction and adaptive learning.
The broader goal is to demonstrate how an agentic operating layer can be extended with new tools, devices, sensors, software, or robotic components, and then adapt its behavior through bounded experimentation and feedback across heterogeneous environments.
For a human, suddenly receiving a new hand or mechanical limb would be profoundly difficult. Even if the limb were physically attached, the brain would still need to construct a new body map, understand the available range of motion, connect intention to movement, and learn how visual feedback corresponds to real control.
Simply knowing that the limb exists would not be sufficient. The person would need to experiment, observe outcomes, make errors, correct them, and gradually develop coordination through continuous feedback.
This experiment applies the same principle within an agentic automation context.
Speedy Evolv is equipped with a new mechanical arm attachment, observes itself through a webcam, and begins mapping actions to physical consequences. The objective is not merely to execute predefined movements, but to progressively form a practical control model through perception, feedback, adaptation, and self-correction.
It is an interesting agentic operating system designed to interface with many kinds of external substrates: software, devices, sensors, robotic attachments, smart TVs, energy systems, switches, meters, and solar infrastructure.
Each external system is managed through a controlled connector layer. If a connector does not yet exist, Speedy Evolv can generate the initial structure for one: it identifies the target substrate, defines the available actions, establishes the control boundary, records the required permissions, and validates the integration through observable readback.
This means the system is not limited to a predefined set of devices. It can extend its operational reach by fabricating new bounded connectors for additional hardware, software, sensors, robotic components, smart-home devices, or energy systems.
he connector enables the agent to discover capabilities, execute bounded actions, observe outcomes, and update its operational context through feedback.
In this experiment, the new mechanical arm is treated like any other extension: a capability that must be discovered, modeled, tested, observed, and progressively integrated into the agent’s operating context.
This lab demo shows Vector acting as a physical endpoint of MD-OS. The system converts a natural-language request into controlled robotic operation: Vector speaks, performs a short movement, stops, and returns operational readback through a bounded connector. The goal is not unchecked autonomy. The goal is to show an operating structure for agents: memory, tools, policies, connectors, execution, verification, and continuity working together as a readable control plane. MD-OS extends natural-language programming from writing software to operating the system itself as an agentic software kernel.
This site shares high-level demonstrations and research notes only. Source code, connector configuration, private runtime state, and internal implementation details are not publicly disclosed at this stage.
