kibbu On-Prem
Self-hosted and air-gapped AI inference.
kibbu On-Prem puts the whole system inside your environment: you run the control plane on your own infrastructure, and your enrolled machines run the models. Connected, it can pull releases and models from kibbu. Air-gapped, nothing leaves your environment: releases and model files arrive as a signed offline bundle you import yourself, and kibbu receives nothing.
Two ways to run it
Connected or air-gapped.
The same control plane and the same releases, set up two ways.
| kibbu On-Prem, connected | kibbu On-Prem, air-gapped | |
|---|---|---|
| Control plane | Run by you, in your datacenter or your own cloud account. | Run by you, with no route to kibbu. |
| What kibbu receives | Nothing about your usage. The control plane sends kibbu no version, usage or machine count. It only downloads releases and model files from kibbu's registry, with your download credential. | Nothing. |
| Releases and models | Pulled from kibbu's registry with your download credential. | Imported from a signed offline bundle, or from your own registry. |
| Network | Outbound to kibbu's registry for releases and models, plus the services you configure. | None outside your environment. Releases and model files arrive as a signed offline bundle or through your own registry. |
What you run
The components.
One control-plane image, installed with a Helm chart on Kubernetes, or with Docker Compose on a single VM for smaller fleets.
| Component | What it is |
|---|---|
| Control plane | The kibbu server: API, admin console, scheduling and the mesh hub. One pod, or a hub plus replicas for larger fleets. |
| PostgreSQL | Your own database. It is required in every deployment; the Compose install can bundle one. |
| Model storage | A built-in store on a 250 GB volume for a single-pod install, or your own S3-compatible bucket for a scaled one. |
| Sign-in | Your identity provider, over OpenID Connect. |
| Your SMTP server, for the emails the console sends. | |
| Mesh endpoint | A static IP and UDP port that your machines reach for the encrypted mesh. |
| The agent | Installed on each machine that serves, through your MDM, as in a hosted deployment. |
External dependencies
Everything it can reach, listed.
The control plane prints this list for your own deployment, and checks each destination, with one command: kibbu-server egress-check.
| Destination | When | Air-gapped |
|---|---|---|
| kibbu's registry | Only if you set a download credential, to pull releases and models. | Refused at startup. |
| Hugging Face | Only if an admin turns on the upstream model fallback, which is off by default. | Always off. |
| Your object storage, identity provider, SMTP and tracing | Only the endpoints you configure. | Your own endpoints, kept inside your environment. |
| A cloud AI provider | Only if you add its key. | Reported by the egress check as a failure. |
There is no license server, no usage reporting and no remote access for kibbu. An air-gapped control plane refuses to start if it is configured with any kibbu destination.
Installation and updates
How it gets in, and how it stays current.
kibbu never pushes into your environment. You decide when each release goes in.
- Install the control plane with the Helm chart or Docker Compose, connect PostgreSQL, storage and your identity provider, and generate its secrets once. Back them up with the database.
- For an air-gapped site, each release comes as a signed offline bundle: container images, the Helm chart, agent installers and the models you chose. You load it with kibbu-server import-bundle.
- Enrolled machines update from your control plane, not from kibbu.
- Self-hosted releases are supported for six months, so you can update on your own schedule.
Models
Models come from kibbu's curated catalog of open models. Hosted, machines download them from kibbu's model mirror and check each file's SHA-256 hash. With kibbu On-Prem, your control plane syncs them from kibbu's registry or imports them from an offline bundle, and machines download from your control plane.
Requirements
What you need in place.
The control plane runs on ordinary infrastructure. The models run on your enrolled machines.
- A Kubernetes cluster, or one Linux VM with Docker Compose for a single-pod install.
- PostgreSQL, and either a 250 GB volume or an S3-compatible bucket for model files.
- An OpenID Connect identity provider.
- A static IP for the mesh that your machines can reach over UDP.
- Machines to serve, from the list below.
| Platform | What serves inference |
|---|---|
| macOS | Apple Silicon (M-series). |
| Windows | x64 with an NVIDIA, AMD or Intel GPU. |
| Linux | x64 with an NVIDIA GPU, on Ubuntu 24.04 or newer. |
| Servers | Linux servers and Kubernetes nodes, grouped as Servers so policies can target them. |
Licensing and support
Agreed with you, in the pilot.
Commercial terms, support and pricing are agreed per pilot.
On-Prem deployments start the same way as hosted ones: a scoping call, then a pilot on one workflow. Tell us whether you need connected or air-gapped from the start, and what your environment allows, and we will plan the install around it.
Keep reading
Start with one workflow.
Tell us about the work you want to keep private, and we will scope a pilot on your own machines.