Run agent factories in production.
runkora is the open source control plane for agent factories. It runs the agent harnesses you already use on your own Kubernetes cluster, and governs everything they do.
Agent harnesses
Claude Code
Codex
OpenCode
01 · Agent execution
How does an agent run?
Every run is durable and keeps its history as a log. Compute comes from a sandbox that is deleted when the run ends.
02 · Identity
Who is the agent?
Every agent has its own identity. A delegated agent works with the access of the person who started the run.
03 · Policy
What stops it from doing damage?
Policy checks every request against grants scoped to a single tool, action or repository. A run stops when it reaches its budget.
04 · Connectivity
What can it talk to?
Agents reach models, MCP servers and external systems through runkora’s gateways without ever holding a credential.
How a run moves through runkora
A run is a single unit of work performed by an agent. It is durable, works in the background, and can be observed by anyone with access while it completes.
Trigger
A trigger starts the run from an event such as a pull request, an alert or a schedule. The API or a person can also start one.
Queue
The run waits as pending until its agent has capacity and then starts its session.
Identity
An autonomous agent’s run acts as the agent. A delegated agent’s run acts as the person who started it and holds no more access than they do.
Sandbox
When the task needs compute, the agent provisions an isolated pod from one of its environments.
Connect
A gateway checks each request against policy and adds the credential for the destination in flight.
Record
Every action traces back to the agent or person that performed it. Model calls are recorded against the run with their cost.
Walkthrough
From an @kora-agent mention to a review
Built in
These parts of the platform ship with runkora and run on your own cluster.
Read the architectureDeploy runkora
runkora installs with Helm onto your own Kubernetes cluster.
01 · Blank cluster
Start from the examples
The reference install brings up SPIRE, Istio and runkora together on an empty cluster.
helmfile -e example apply02 · SPIRE and Istio already running
Install with Helm
Install runkora onto your mesh with your own values.
helm upgrade --install kora \
oci://ghcr.io/runkora/charts/kora \
-n kora-system -f kora.yaml03 · GitOps
Sync with Argo CD
Apply the same releases as an app-of-apps from your own GitOps repository.
- repoURL: ghcr.io/runkora/charts
chart: kora
helm:
valueFiles: [$values/kora.yaml]Questions
Is runkora open source?
Yes, runkora is licensed under Apache 2.0.
Where does runkora run?
runkora runs on your own Kubernetes 1.34 or later cluster with Istio and SPIRE installed.
Which agent harnesses does it run?
runkora runs Claude Code, Codex and OpenCode, and starts each harness with the agent’s model, tools, environments and identity.
Which models can agents use?
Agents can use any model from a provider that the platform team registers with the inference gateway. Each agent calls a virtual model by its alias, and the platform team can move the alias to another model or provider.
Do agents ever see credentials?
No. The inference and egress gateways inject credentials into requests in flight, and each MCP server holds its own. A sandbox holds no credentials.
Is runkora an agent framework?
No. runkora runs the harness an agent was built for and governs what the agent does.