Skip to content

Repository files navigation

Kaimahi night worker guarding paths for AI agents

Kaimahi

Get agents onto Orka

Orka is the platform. Kaimahi is incubating tooling that helps people get agents onto it, with particular attention to Kubernetes and AKS. It is not another agent platform.

kmx prepares a cluster, installs a pinned Orka, reports what is actually running, authors native Orka Agents, and helps route an existing application's model traffic through Orka. The application's owner keeps its Deployment and lifecycle. The seam between the application and Orka is a bridge: shrinking it to nothing is success, not lost product scope.

Quickstart

The Orka helpers are on main; the latest tagged release, v0.1.0, predates them. For this development path, install Go 1.26+ and Docker or Podman, ensure your Go binary directory is on PATH, then:

go install github.com/kaimahi-agents/kaimahi/cmd/kmx@main
kmx up
kmx orka install
kmx orka status

@main follows a moving development branch, not a stable release. Use a reviewed commit instead when you need a reproducible CLI build.

kmx up currently creates a local kind cluster, Ollama and the existing kagent runtime. It is not an Orka-native setup command. kmx orka install then installs Orka's pinned manifest and a keyless local Provider; it does not migrate or govern an application. kmx orka status distinguishes the running controller version from the version kmx pins.

Already have a cluster? Read getting started and the Orka installer contract, including target confirmation, --no-apply, --dry-run, and the Provider prerequisites. For cloud setup, read AKS before creating billable resources.

From a checkout, make builds bin/kmx; use that binary for the same commands. Installation and releases describes tagged binaries, checksums, and upgrade limits.

Migrate model traffic

For an application already deployed and managed by its owner, the current bridge uses the retained plane implementation:

kmx plane
kmx migrate <deployment> --namespace <namespace> --model <provider>/<model>

The placeholders must name your existing workload and an Orka Provider. Read the migration guide before running this: inspect and apply the workload patch the command prints. kmx migrate does not patch your Deployment for you, adopt it, convert it into an Orka Agent, or create Orka Tasks for its requests.

Governance here means governed model traffic, not application ownership or automatic governance of every tool, network connection, and inbound event. Authentication, Provider scope, recording, protocol translation, credential renewal and known limitations are documented at the migration boundary. Installing Orka alone enables none of this routing.

Status

  • Current tooling: Orka installation/status, native kmx agent create (Provider + Agent, optionally a Task with an actual answer), and model-traffic migration. See the native create guide. Migration was exercised on kind and AKS; cloud runs are measurements, not a continuously maintained deployment. See migration.
  • Authoring is open: whether the supported authoring surface will be native Orka only or also kagent YAML over Orka is not decided. docs/orka.md recommends native resources; that is a recommendation, not a ruling. Today's kmx agent create emits native Orka resources; it does not convert kagent YAML, ModelConfigs, MCP wiring or BYO images. The isolated conversion spike is not a supported CLI interface.
  • Legacy implementation remains: the governance plane, its connectors, demonstrations and kagent commands still exist. Their documentation is labelled as legacy reference, not a list of Orka capabilities or a promise that Kaimahi will continue as a platform. Code retirement is a separate decision and lane.
  • Upstream first: do not rebuild what Orka supplies. orka.harness.v2 is not a direction for this project. OTLP with GenAI conventions ships in Orka; it is not an outstanding Kaimahi upstream candidate.

Documentation

Start at the documentation index, which separates current operator paths from references for the legacy code still in this tree.

Development

Read CONTRIBUTING.md and the entry-point principles. Changes land via pull requests to main with checks green and verification actually run. The project name's cultural and publication boundaries remain in docs/NAMING.md.

About

Build and govern cloud-native AI agents on Kubernetes. kmx is the developer entry point; budgets, credential custody, tool controls, audit and bounded approvals come built in.

Topics

Resources

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages