DiMOS open‑sources agent OS
A social post highlights that DiMOS has open‑sourced an agent‑native operating system designed to let large language models control quadrupeds, humanoids and drones in real time. The release claims to simplify robotics stacks by providing a runtime that connects LLM decision-making to low‑level control pathways. (x.com)
Robots already run on layers of software that turn camera feeds and sensor data into motor commands. DiMOS says it has open-sourced one stack meant to let language-model agents sit inside that loop. (github.com) The code is public on GitHub under the `dimensionalOS/dimos` repository, which showed about 2,500 stars and nearly 400 forks when crawled this week. The project page calls it “the agentic operating system for physical space.” (github.com) A March 11, 2026 package release on Python Package Index and a March 12, 2026 GitHub release marked version 0.0.11. The release notes said the update added “first-class support for coding agents,” Model Context Protocol tooling, drone support, temporal memory and fleet control. (pypi.org) (github.com) In plain terms, the software tries to do for robots what an operating system does for a laptop: handle the plumbing between applications and hardware. DiMOS says its modules pass typed data streams over transports including Lightweight Communications and Marshalling, Robot Operating System 2 and Data Distribution Service, while “blueprints” bundle those pieces into runnable robot stacks. (github.com) The pitch lands at a moment when robotics developers are trying to connect large language models to machines that must react in milliseconds, not just answer text prompts. DiMOS says its agents run as native modules that can subscribe directly to perception, memory and control-loop streams instead of sitting outside the robot as a separate chat layer. (pypi.org) The project also leans on a simpler developer workflow than many robotics stacks. Its package page says developers can build applications entirely in Python with “no ROS required,” even though the framework can still communicate over Robot Operating System 2 when needed. (pypi.org) (github.com) The current docs list example blueprints for a Unitree Go2 quadruped, a Unitree G1 humanoid, xArm manipulators and drone replays. One quick-start section shows Go2 and G1 runs in replay, simulation and real-hardware modes, while the drone module describes DJI integration through a RosettaDrone MAVLink bridge. (github.com 1) (github.com 2) Not every agent feature is available on every robot yet. The AGENTS guide says the only shipped blueprint that currently exposes a live Model Context Protocol server is `unitree-go2-agentic-mcp`, while other agentic blueprints use an in-process agent module instead. (github.com) The maintainers also describe the project as a “Pre-Release Beta,” which is a reminder that open source does not mean production-ready by default. For now, DiMOS is offering a public codebase, a Python package and a claim that one runtime can span quadrupeds, humanoids and drones without a separate stack for each. (pypi.org) (github.com)
