What is Orka?

Orka runs AI work on Kubernetes. You describe a piece of work as a Task — answer a question, fix a bug and open a pull request, review every pull request in a repository — and Orka runs it in your cluster, keeps a durable record of what happened, and returns the result through a REST API, a CLI, a web dashboard, or the chat and editor tools your team already uses.

Model API keys and Git credentials stay in the cluster. People, CI jobs, and agents get scoped client credentials — Kubernetes ServiceAccount tokens, OIDC tokens, or transaction tokens — instead, and the platform team decides which models, tools, and repositories are allowed.

See it before you install it

The dashboard demo runs the real Orka dashboard in your browser on sample data for a fictional company, with a guided tour. Nothing is sent to a cluster or a model. The demo videos show the same features on a live cluster.

Experimental

Orka is under active development. APIs, custom resources, and behavior may change between releases, and it is not yet recommended for production use.

How it works

Three custom resources cover most of what you write:

Resource	What it is
Provider	A model backend — Anthropic, OpenAI, Azure OpenAI, or any compatible endpoint — and the Secret that holds its API key.
Agent	A reusable configuration: model, system prompt, tools, and optionally a coding-agent runtime. A template, not a running process.
Task	One unit of work. Create a Task to make something happen.

A Task points at an Agent, and an Agent points at a Provider. A Task has one of three types, and each type runs somewhere different:

Type	What runs	Where it runs
ai	Orka's own AI worker, which calls the model in a loop with built-in tools such as web search and code execution	A per-Task Kubernetes Job
agent	A coding-agent CLI — Codex, Claude Code, GitHub Copilot CLI, or OpenCode — working on a git repository	A warm, controller-owned pool, so later Tasks start in seconds
container	Any container command. No model involved.	A per-Task Kubernetes Job

Every Task leaves a record you can come back to: its phase, result, artifacts, logs, and an event timeline. A coding agent never holds a Git credential. When it changes a repository, a separate publisher pushes the change and Orka records a verified delivery receipt for the branch or pull request. Architecture explains the full picture, and the Glossary defines every term.

What you can do with Orka

Run AI work

• Ask a model, with built-in tools for web search, code execution, URL fetching, and file access. Getting started
• Hand a coding task to an agent. Codex, Claude Code, Copilot CLI, or OpenCode works on a clean checkout, and Orka publishes the branch and pull request. Agent runtimes
• Run build and test steps in any container image. Container tasks
• Run any Task on a schedule, with missed-run and overlap rules. Scheduled tasks
• Continue a conversation across Tasks with Sessions. Session continuity

Let agents work together

• Delegate. A coordinator agent splits work across specialists, within the depth, allowlist, and concurrency limits you set. Multi-agent coordination
• Work toward a goal. An autonomous coordinator plans, delegates, and re-plans across iterations until the goal is met. Autonomous tasks
• Remember. Agents propose memories, and people review them before any agent uses them. Memory

Automate your repositories

• Review every pull request as it arrives and post findings back. Repository monitors
• Turn an issue into a pull request by adding the orka:implement label. Planning, implementation, review, and repair follow. Issue-to-PR automation
• Find security issues with a threat model, validated findings, and remediation pull requests. Security scanning
• Measure the cost of delivery as tokens per merged pull request. Usage and PR outcomes

Meet people where they work

• Chat with an orchestrator that creates and runs Tasks for you. Interactive chat
• Watch and steer work in the web dashboard. Web dashboard
• Script it with the orka CLI. CLI reference
• Point Claude Code, Cursor, or Continue at Orka through OpenAI- and Anthropic-compatible endpoints, so nobody needs a provider key. OpenAI-compatible API · Anthropic-compatible API
• Reach agents from Microsoft Teams, Telegram, or other agents over A2A. Gateways

Stay in control

• Keep credentials away from coding agents. Their model keys sit behind a proxy, and Git credentials go only to the publisher. Security
• Approve consequential actions before they run. Human approval
• Give agents your own tools — HTTP APIs and MCP servers — behind outbound access policies. Agents, skills, and tools
• Let people act as themselves by linking their own GitHub account, with approval on every write. Connectors
• Trace everything with OpenTelemetry and a durable per-Task event history. Observability · Execution events

Extend Orka

• Bring your own agent runtime, such as an AgentKit container or a Microsoft Foundry hosted agent. Bring your own agent runtime
• Run agents in sandboxes that suspend, resume, and checkpoint their files. Execution workspaces

Find your way

If you want to...	Start here
Try Orka on a laptop	Install Orka, then Getting started
See it working first	Dashboard demo and Demos
Copy a working setup	Examples on GitHub
Understand how it fits together	Architecture and Glossary
Run coding agents with your models	Provider proxy, then Agent runtimes
Run it for a team	Security, Configuration, and Upgrading
Fix something	Troubleshooting and the Operations runbook
Build on the API	API reference and API authorization
Contribute to Orka	Development

For agents and automation

If you are an AI agent or a script learning Orka, these pages are the dense ones:

• Glossary for the vocabulary, and Architecture for components and lifecycles.
• Resource reference for every custom resource field.
• API reference and API authorization for HTTP routes and the permissions each needs.
• CLI command reference for every command and flag, generated from the binary.
• Built-in tools for the tools Orka gives agents and their parameters.

Two rules prevent most surprises. Create resources in the namespace the controller watches (orka-system by default), because resources anywhere else are silently ignored. And treat a Task as done only when its phase is Succeeded — for a coding agent that changes a repository, status.delivery also records the verified result. To work on the Orka source itself, read AGENTS.md in the repository.

Ecosystem

These projects live outside the main repository and are versioned separately:

Project	What it adds
orka-gateway-teams	Talk to Orka Agents from Microsoft Teams chats and channels.
orka-gateway-telegram	Talk to Orka Agents from Telegram private chats.
orka-gateway-a2a	Call Orka Agents from any A2A client.
agentkit	Build an agent from YAML into a container image, then run it as an Orka runtime.
orka-runtime-foundry	Run Microsoft Foundry hosted agents as Orka runtimes.
orka-integration-kontxt	Issue and exchange transaction tokens with Kontxt.
orka-integration-agentgateway	Route outbound tool calls through agentgateway.
orka-workspace	The shared execution workspace API and its provider controllers.
sessionkit	Move Codex conversations between machines and Orka.
Vekil	A model gateway that fronts GitHub Copilot, OpenAI, Anthropic, and others for coding agents.

Get involved

Questions, bug reports, and ideas are welcome in GitHub issues. To contribute code, start with CONTRIBUTING.md and the Development guide. Report vulnerabilities privately as described in SECURITY.md.