Skip to main content

Open-source agent runtime

The operating system for enterprise AI agents

Open Agent Kernel (OAK) is the open-source runtime for scalable, compliant agents. Build in the framework you already use and run on any cloud.

pip install agentkernel

Using a coding assistant? Install the OAK agent skills

01What OAK does

Six stages every agent request passes through

Each one is switched on with a few lines of configuration, not built from scratch.

Build

Write agents in the framework you already know. OAK wraps them with one API, so the agent logic stays yours.

Learn more
  • Adapters for eight agent frameworks, side by side in one runtime
  • Tools written once as plain Python and bound to any framework
  • Structured output, multimodal input and multi-agent handoffs
my_agent.py
from agents import Agent as OpenAIAgent
from agentkernel.cli import CLI
from agentkernel.openai import OpenAIModule

assistant = OpenAIAgent(
name="assistant",
instructions="Give short and direct answers.",
)

# Register the agent with OAK
OpenAIModule([assistant])

if __name__ == "__main__":
CLI.main()

02Architecture

One runtime between your channels and your agents

Requests arrive from any channel or protocol. OAK runs your framework-native agents with sessions, guardrails, queues and sandboxes, on whichever cloud you choose.

Sources · channels and clients

Messaging channels

Protocols and APIs

Apps, voice and schedules

Destinations · agents, data and clouds

Agent frameworks

Memory and knowledge

Clouds and observability

Request inBuildConnectRememberGuardScaleObserveResponse out

03Works with

Plug in what you already run

75 integrations, and every one is an adapter. Swap one without touching the others.

Read the integration guides

04Deployment

Deploy anywhere

Run the same agent code on AWS, Azure, GCP or your own Kubernetes cluster. Terraform modules and a Helm chart are included.

Amazon AWS

Serverless or containerized deployments with Terraform modules.

  • AWS Lambda (serverless)
  • AWS ECS on Fargate (containerized)

Microsoft Azure

Container Apps with Cosmos DB session storage.

  • Azure Container Apps (containerized)
  • Queue mode over Azure Service Bus

Google Cloud

Cloud Run serverless or containerized deployments with Firestore session storage.

  • Cloud Run (serverless)
  • Cloud Run (containerized)

On-prem Kubernetes

The official Helm chart, for any Kubernetes cluster.

  • Bare metal or a self-hosted cluster
  • AWS EKS (Pod Identity, with SQS, NATS or Kafka)
  • KEDA autoscaling on queue depth
  • Air-gapped installs (mirrored images)

06Questions

Questions we hear a lot

What OAK is, what it costs, and how it fits with the tools you already use.

Ship your first agent today

Free and open source under Apache 2.0. No licensing costs and no lock-in.