Orka
A Kubernetes-native orchestration platform for AI agents: a YAML Task runs a model call, a coding agent such as Codex or Claude Code, or a plain container command in an isolated pod, with the model keys staying inside the cluster.
Orka lets a user hand a piece of work to an AI agent and get the result back, with Kubernetes running the work in an isolated pod, keeping a record of what happened, and cleaning up afterward. A small YAML Task describes one of three kinds of work: asking a model directly through Orka's own AI worker, which talks to Anthropic, OpenAI, Azure OpenAI or a compatible endpoint with built-in tools such as web search and code execution; running a coding agent such as Codex, Claude Code, GitHub Copilot CLI or OpenCode against a real git repository, opening a pull request when it finishes; or running an arbitrary container image for build and test steps an agent needs.
Results are read back from a REST API, a CLI, or a built-in dashboard, and model API keys never leave the cluster: people and CI get a token, while the platform team decides which models and providers are allowed. Installation is a single Helm chart against an existing Kubernetes cluster, including a local kind cluster for trying it out with no model key required at all.
It is written in Go, described as experimental with no production guarantee yet, and governed under Microsoft policy with a Microsoft contributor agreement required from external contributors pending a planned donation to a community foundation.