Model or dataset
OpenMind/OM1 avatar
OpenMind/OM1

OM1: the container asks for host networking, two capabilities and no AppArmor

Modular AI HAL (Hardware Abstraction Layer) for Robots

2,967 stars996 forksGoMIT

At a glance

What is it?
A Go runtime for multimodal agents that drive humanoids, phones and quadrupeds, speaking to robot middleware through plugins. It is MIT licensed but needs a commercial API key and bills in credits. The repository ships a compose file that runs the container on the host network with two Linux capabilities added and the mandatory access control turned off.
Who is it for?
Adopt this if you are building on real robot hardware and want one runtime that spans a laptop, a phone and a humanoid, and read the deployment section before you run it on a machine with anything else on it. Three things to weigh.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
No. The owners have archived the repository on GitHub, so it is read-only and no longer receives changes.
What is it written in?
Mainly Go, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • OpenMind/OM12,967 ★ · +60
Star history
Repository2026-09-102026-11-15
OpenMind/OM12.9k3k

60 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The compose file runs the container on the host network with two capabilities

The deployment configuration is where the substance of this project is, and it is worth reading as a security decision rather than as packaging. The main service sets its network mode to host, so the container is not on a bridge network and shares the host's interfaces. It then adds two Linux capabilities, one for tracing processes and one for network administration, disables AppArmor entirely through a security option, and passes two host devices straight through, the sound device and the first video device. It also joins the audio group and maps several directories out of the user's home into the container, two of them writable. None of that is wrong for a robot that needs to reach a sound card and a camera, and all of it means a compromise of the process is a compromise of the host. A second service in the same file, a local model server, is also on the host network and asks for the NVIDIA runtime.

The module asks for one language version and the image ships another

Three places state the toolchain version and they do not agree. The module file asks for a language version of one point twenty five point zero. The readme's build-from-source section says the same, one point twenty five point zero or newer. The container image's builder stage is a different line again, a Go one point twenty six image on a Debian base. So the container compiles the code with a newer compiler than the one the module declares as its floor, and a developer following the readme will build with a version one release behind the image. The final image stage is a slim Debian, not the same base as the builder, and it adds a long list of audio libraries, rewrites the sound configuration to route through PulseAudio, and patches the name resolution file to add minimal multicast DNS.

The module path and the repository name use different capitalisation

The repository is named with a capital letter on both segments. The module file declares its path in all lower case. The clone command in the readme uses the capitalised form of the repository. The published container image uses a third spelling, with a different organisation segment and a lower case project name, and the container itself is named with a short fourth form. That is four identifiers for one project, which is the kind of drift that makes a link break or a search fail. The go file inside the image is consistent with the capitalised repository form, and the module path is the one that is not, so a tool reading the module metadata and a tool reading the repository will not match them.

The newest release is a tag called nightly and it is four months old

There are three releases and the newest one is not a version. It is tagged with the single word nightly and titled as a development build, published on 2026-06-04. The two before it are beta tags of the same one point zero point two line, fifteen days apart, from mid and late April 2026. The last push to the default branch is dated 2026-09-30, roughly four months after the nightly tag was cut, so the source has moved substantially since anything downloadable was published. There is no stable tag in the visible history, and a tag whose name is nightly carries no information about which commit it contains, so a script pinning it gets a different binary on every run.

Four pre-built platforms and none of them is Windows

The download section lists four pre-built targets, and the readme says the pre-built route needs no Go installation. Those four are Linux on two architectures and macOS on two. Windows is not among them, and the file does not say so; a reader scanning for their platform finds a list and an absence rather than a statement. The archive is described as already containing the configuration directory, the knowledge base and the middleware library, so a user does not need to clone the repository for runtime assets. The build-from-source alternative needs the compiler and the make tool, runs a dependency step before a build step, and downloads two further artefacts, a local inference runtime and a voice activity detection model.

On macOS the binary is quarantined and its library is not on the default path

Two separate macOS steps, and the second one is the interesting one. The first clears the quarantine attribute that makes a freshly downloaded unsigned binary refuse to run, and the readme offers it as a command that suppresses its own errors. The second sets a dynamic library path variable to the current directory before launching:

bash
export DYLD_LIBRARY_PATH="$PWD:$DYLD_LIBRARY_PATH"
./om1 -config ./config/conversation.json5

The Linux equivalent sets the other spelling of the same variable. So the promise of a single self-contained binary is not quite delivered on either desktop platform: the program finds the bundled middleware library by being told to look in the directory it was launched from. The readme also tells you to grant the binary, and the middleware library if prompted, permission in the system privacy settings, and the container route makes the same request in the form of a device passthrough.

The runtime needs a commercial key and bills in a unit of its own

The source is MIT licensed, and the runtime is not free to run. A key is required before the agent will start, obtained from a vendor portal, and it can be supplied through the shell profile or through a project-local environment file copied from an example. Billing is in a unit the project calls OMCU, described as the computational unit for billing on the platform's plan, and the free plan is fifty of them renewed monthly, with paid upgrades sold through the same portal. The troubleshooting section's first item is an authentication error, which follows from that. So the dependency chain for running this is a local binary, a middleware library, and a metered remote service, and only the first two are in the repository.

The architecture section is an empty heading and the closing list stops mid-sentence

Two small gaps in a long readme, both easy to confirm. The architecture overview section consists of a heading followed immediately by the getting started section, with nothing between them, so the one section that would explain how the pipeline is put together is the one section with no content. And the closing list of what to try next ends on a bullet that begins by telling you to change the system prompts in the configuration and then stops. Neither is a functional problem, and both are worth knowing about before you conclude the readme describes the whole system. What it does document in depth is the install and run path, which is unusually complete for a project of this size.

Editorial conclusion

Adopt this if you are building on real robot hardware and want one runtime that spans a laptop, a phone and a humanoid, and read the deployment section before you run it on a machine with anything else on it. Three things to weigh. The compose configuration is the most privileged thing in the repository, so treat it as a starting point to narrow rather than a reference to copy. The newest release tag is the word nightly rather than a version, so there is no meaningful version to pin. And the runtime bills against a commercial service, which is worth weighing against the permissive source licence before you build a product on it.

Frequently asked questions

What is OpenMind/OM1?

It is a modular runtime for multimodal AI agents, written in Go, intended for robots and other physical or simulated environments. It handles inputs such as web data, camera feeds and sensor data, and actions such as motion, navigation and speech, and is configured with json5 files that combine inputs and actions.

Does OM1 need a paid service to run?

It needs an API key from the vendor's portal. Billing is in a unit called OMCU, described as the computational unit for the platform plan, and the free plan provides fifty of them renewed monthly. The source itself is MIT licensed.

Which robot middleware does OM1 support?

Hardware support is provided through plugins that reach three middleware systems, and the readme recommends one of them for all new development. The three named are a robotics middleware, a pub and sub middleware, and a data distribution service, and the dependency list includes the Go binding for the recommended one.

Can I build and run OM1 on Windows?

The readme does not address Windows. The pre-built binaries are listed for two Linux architectures and two macOS architectures, and the build-from-source instructions cover a Unix-style path with make. Windows is an absence in the list rather than a stated exclusion.

What does the OM1 docker compose file request?

Host network mode rather than a bridge network, two added Linux capabilities for process tracing and network administration, AppArmor disabled, the sound and video devices passed through, and several directories mapped from the user's home into the container, two of them writable. A second service for local model serving is also on the host network and requests the GPU runtime.

Official sources

  1. License: MIT
  2. OpenMind/OM1 on GitHub
  3. Project website
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/openmind-om1.svg)](https://hysenlabs.com/projects/openmind-om1)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/openmind-om1.svg)](https://hysenlabs.com/projects/openmind-om1)

Was this useful?