Kubemoot

Every voice, one answer.

Declare a crew of AI agents on Kubernetes. Kubemoot convenes them, schedules their models on the accelerators you choose, governs what they say, and lets them deliberate to one answer you can measure.

Install Quickstart View on GitHub

A moot is an assembly that settles a question by deliberation. Kubemoot runs a crew of AI agents on Kubernetes that each bring their own expertise and reach consensus over a message bus, where any of them can agree, raise a concern, object, or stand aside. Capability is composed horizontally from many small models rather than concentrated in one frontier model. Kubemoot focuses on crews of agents reaching consensus on shared accelerators, and runs alongside projects such as kagent and agentgateway.

Consensus, not one model

Specialists deliberate and settle by signal - no single model, no fixed pipeline, no majority vote. Failure is first-class signal, not silence.

Declarative crews

Crews, agents, prompts (ADL), and scheduling policies are Kubernetes objects. Author by kubectl apply; version with git diff; no recompilation.

Runs on your accelerators

Just-in-time model scheduling spreads small models across commodity or local accelerators. A crew states the models it needs and the cluster’s catalog supplies them, so a crew is portable across clusters; no frontier-model dependency.

Measurable

Executable fitness functions score a crew against ground truth - reference-grounded, fabrication-aware - so crew quality is something you track, not guess.

Built on open source, swappable

Tools through MCP, the bus on NATS, models on vLLM or Ollama: each sits behind a seam you can swap. Kubemoot builds the moot and convenes the rest.

What Kubemoot stands on

You declare the crew and how it decides; Kubemoot runs the discussion. Each of these sits behind a swappable seam:

  • Tools through MCP.
  • The message bus on NATS.
  • Model servers: vLLM by default, and Ollama, through the Model Server Protocol.
  • Isolation: Kubernetes RBAC and network policies, with code run in a sandboxed pod.
  • Other harnesses at the table: goose already holds a seat.
  • Crew memory exported in the Open Memory Protocol (OMPI) format.

See What it builds, what it convenes.

Five-minute start

A CPU trial you can run on a laptop, no accelerator required: one script takes an empty cluster to a crew answering a question. NATS, Ollama serving a small model from the cluster’s model catalog on CPU, the operator, and a two-agent crew.

git clone https://github.com/kubemoot/kubemoot.git && cd kubemoot
git checkout "$(git describe --tags --abbrev=0 --match 'operator-chart-v[0-9]*' --exclude '*-rc.*')"
kind create cluster --name kubemoot
./quickstart/quickstart.sh

The git checkout selects the latest release, and the script installs the chart of the release it is checked out at. The same script runs in CI against every release, on a 2-CPU node. What each step creates, the CPU-trial-versus-accelerator profile, and how to go further: Quickstart. To install the operator on your own cluster with an accelerator-backed model server (vLLM is the preferred engine): Installation.

The ecosystem

  • Kubemoot - the operator: the Kubernetes controller and runtime that wire a crew together.
  • Crews - packaged crews as Helm charts, from the reference crew to examples you can copy.
  • Homelab Pilot - the reference crew, with its chat and dashboard app.
  • kmctl - the command-line tool for crews, discussions, and the cluster’s model catalog.
  • CrewForge - the VS Code extension for building, deploying, and talking to crews.

See Ecosystem for what each component is and how they fit.

Community

Kubemoot is an independent open-source project under the Apache 2.0 license. Start with the Community section and the Contributing page, the kubemoot organization on GitHub, and Discussions. For anything else, write to moot@kubemoot.org. security@kubemoot.org is only for vulnerability reports.

OpenSSF Best Practices OpenSSF Scorecard

See How Kubemoot is built and secured.