Orka 发布：在 Kubernetes 上运行 AI 智能体任务

Orka lets you hand a piece of work to an AI agent and get the result back, with Kubernetes doing what it is good at: running the work somewhere isolated, keeping a record of what happened, and cleaning up afterwards. You write a small YAML Task, Orka runs it in a Pod, and you read the answer from a REST API, a CLI, or the dashboard that comes built in.

Three kinds of work fit in a Task:

Ask a model. Orka's own AI worker talks to Anthropic, OpenAI, Azure OpenAI, or any compatible endpoint, with built-in tools such as web search and code execution.
Run a coding agent. Codex, Claude Code, GitHub Copilot CLI, or OpenCode, working on a real git repository and opening the pull request when it is done.
Run a command. Any container image, for the build and test steps agents ask for.

Model keys never leave the cluster. People and CI get a token; the platform team decides which models and providers are allowed.

Important

Orka is experimental and under active development. APIs, CRDs, and behavior may change without notice between releases, and it is not yet recommended for production use. Feedback, bug reports, and feature ideas are very welcome — please open an issue.

Note

The organization and repositories are intended to be donated to a community-governed foundation at the appropriate time. Until then, the project is governed by Microsoft policy, and external contributors are required to sign the Microsoft Contributor License Agreement (CLA).

Try it in five minutes

You need a cluster, kubectl, and Helm. A local kind cluster is fine. No model API key yet.

helm repo add orka https://orka-agents.github.io/orka/charts
helm repo update orka
helm install orka orka/orka --namespace orka-system --create-namespace --wait --timeout 10m

That is the whole install. Now run something:

kubectl -n orka-system apply -f - <<'EOF'
apiVersion: core.orka.ai/v1alpha1
kind: Task
metadata:
  name: hello
spec:
  type: container
  command: ["sh", "-c", "echo hello from orka"]
EOF
kubectl -n orka-system get task hello --watch

When the phase reads Succeeded, Orka has run its first Task for you. From here, Getting started walks through connecting to the API, adding a model, and running your first AI Task and your first coding agent.

What people build with it

Pull request review on every PR. A repository monitor queues review Tasks as PRs arrive and posts the findings back.
"Just do it" chat. Describe what you want in plain language and an orchestrator creates and runs the Tasks, in the dashboard or over the API.
Agents on a schedule. Nightly dependency audits, security scans, or reports, with retries and notifications handled for you.
Your existing tools, safer. Point Cursor, Continue, or Claude Code at Orka's OpenAI- and Anthropic-compatible endpoints and stop handing out provider keys.

Why on Kubernetes

No keys on laptops. Provider credentials live in Secrets; people and CI get scoped tokens.
One place to govern. Models, providers, tools, and limits are set per Agent and per namespace.
Everything is a record. Tasks, sessions, artifacts, and the pull requests agents open are durable and auditable.
Scale with the cluster. Scheduling, retries, concurrency, and cron come from the control plane you already run.

Learn more

Architecture — how a Task becomes a Pod
Interactive chat — the orchestrator and its tools
Coding agents — Codex, Claude Code, Copilot, and OpenCode as pooled runtimes
Compatibility APIs — using Orka from your editor
Security — the trust model and hardening
Troubleshooting — error strings, causes, fixes
Development — building and testing Orka itself

Join in

Questions, bug reports, and ideas are welcome in issues. If you want to contribute code, CONTRIBUTING.md has the setup and the review process, and SECURITY.md explains how to report a vulnerability privately.

License

MIT