Deployment
One platform. Three ways to run it.
The same governed AI coding platform, deployed where your data rules require it. Start on our cloud. Move inference into your network. Or run the whole thing inside your perimeter, air-gapped.
Kimchi Cloud
We run the control plane, the gateway, and the models. You call an endpoint.
Your inference, our control plane
The gateway and models live in your network. We run the console and governance.
The full stack, inside
The whole platform in your cluster. One Helm chart. Air-gapped is this, with reporting off.
Your execution, our models
The gateway and remote workspaces run in your network. Inference is served by Kimchi.
What runs where. What leaves your network.
The same policy engine in every pattern. What changes is who operates it and which bytes cross the boundary.
| Component | Kimchi Cloud | Hybrid | Self-hosted |
|---|---|---|---|
| Control plane | Kimchi cloud | Kimchi cloud | Your cluster |
| LLM gateway | Kimchi cloud | Your cluster | Your cluster |
| Inference | Kimchi cloudOpen models on our GPUs. | Your clusterYour models, or Kimchi serverless as overflow. | Your clusterYour models. Any in-perimeter endpoint. |
| Console | Kimchi cloud | Kimchi cloud | Your cluster |
| Analytics store | Kimchi cloud | Kimchi cloud | Your cluster |
| Identity | Kimchi cloud | Kimchi cloud | Your IdPOIDC. Nothing to provision on our side. |
| Leaves your network | Prompts and code | Minimal operational metadata | Nothing, with reporting disabledOptional reporting, off in the air-gapped configuration. |
Runs inside. Answers to your team.
Self-hosted puts the whole platform on your infrastructure, under your identity, behind your monitoring. No dependency on a round trip to ours.
OIDC login, your IdP
Login runs through your OIDC provider. Roles, bindings, and API keys live in the identity services inside your cluster. No Cast AI account required.
Governance that doesn't phone home
Budgets, usage analytics, and the per-request audit trail are stored in the databases that ship with the stack. They keep working with no outbound connection.
Your inference, fronted
The gateway sits in front of what you already run - a vLLM server, an existing endpoint. Models you point it at are the models it serves.
Start where it's easy. Move when the rules change.
The API is the same in every pattern. Migration is a config change, not a rewrite.
Cloud to hybrid
Register your inference endpoints. Policy keeps deciding which requests stay in and which can still use serverless. Developers keep the same SDK. Only the endpoint may change.
Hybrid to self-hosted
Install the stack on your cluster, point clients at the internal endpoint. Budgets, analytics, and the audit trail come with it.
Self-hosted to air-gapped
Restrict inference to in-perimeter endpoints, isolate the network. That's the whole step.