⚒️ ark
Agentic Runtime for Kubernetes
A declarative, Kubernetes-native framework for building portable, scalable, and provider-agnostic agentic applications.
Features • Quickstart • Documentation
What is Ark?
Ark is a declarative toolkit for building and hosting distributed AI agents. By defining what agents should do rather than how they do it, Ark eliminates vendor lock-in and ensures your applications stay adaptable as AI evolves.
Built on Kubernetes, Ark lets you deploy a dev-friendly cluster in minutes or scale agentic workloads across existing infrastructure. Leverage proven patterns for security, monitoring, and RBAC—avoiding bespoke overhead while maintaining a portable, production-ready foundation for your AI projects.
Features
- Declarative Agents — Define agents as Kubernetes custom resources with prompts, tools, and model references
- Provider Agnostic — Swap between OpenAI, Anthropic, Google, Azure, or local Ollama without code changes
- Multi-Agent Teams — Orchestrate agents with sequential, graph, selector, or round-robin strategies
- Tool Integration — Connect agents to HTTP APIs, MCP servers, or other agents as tools
- Persistent Memory — Maintain conversation context across sessions with pluggable memory backends
- A2A Protocol — Interoperate with external agent systems via Agent-to-Agent protocol
- CLI & SDKs — Manage agents from the command line or integrate via Python and TypeScript SDKs
Quickstart
You will need a Kubernetes cluster to install Ark into. You can use Minikube, Kind, Docker Desktop or similar to run a local cluster.
Ensure you have Node.js and Helm installed. Then run the following commands to install Ark:
# Install the 'ark' CLI:
npm install -g @agents-at-scale/ark
# Install Ark:
ark install
# Optionally configure a 'default' model to use for agents:
ark models create default
# Run the dashboard:
ark dashboard
In most cases the default installation options will be sufficient. This will install the Ark dependencies, the controller, the APIs and the dashboard. You can optionally setup a default model that will be the default used by agents. The install command will warn if any required dependencies are missing.
User guides, developer guides, operations guides and API reference documentation is all available at:
https://mckinsey.github.io/agents-at-scale-ark/
To troubleshoot an installation, run ark status.
Credits
The initial design and implementation of Ark was led by Roman Galeev, Dave Kerr, and Chris Madden.