# MCP server

> Ownkube runs a hosted remote MCP server. Connect any MCP client over OAuth and let your AI agents inspect clusters, ship deployments, and manage environments.

Ownkube runs a hosted **remote MCP server**. Point any MCP-capable client at one URL, sign in once, and your agent can inspect clusters, ship deployments, tail activity, and manage environments on your behalf. It uses the same permissions and tenant isolation as your dashboard account, and Ownkube never receives your cloud keys.

MCP (the Model Context Protocol) is an open standard for connecting agents to tools. Because this is a remote server that speaks OAuth, any client with remote-MCP support connects the same way: Claude, other desktop and web assistants, agent frameworks, and your own code via an MCP SDK. The walkthrough below uses Claude as a concrete example; the URL and the flow are identical everywhere else.

## The server URL

```
https://app.ownkube.io/api/mcp
```

That is the only value you need. Authentication happens in the browser: there is no token or API key to copy, and nothing secret to paste into a config file.

## Connect a client

These steps use Claude. Other MCP clients follow the same shape: add a remote server, authorize in the browser, pick an organization.

1. **Add the connector**

   In Claude, open **Settings > Connectors**, choose **Add custom connector**, and paste the server URL:

   ```
   https://app.ownkube.io/api/mcp
   ```

2. **Sign in to Ownkube**

   Claude opens your browser to authorize. Sign in to your Ownkube account the same way you sign in to the dashboard.

3. **Pick an organization**

   If your account belongs to more than one organization, an approval screen asks which one this connection should act on. Choose it, then approve. The connection is bound to that organization until you remove it.

4. **Start driving Ownkube**

   Back in Claude, ask for what you need, for example *"list my Ownkube deployments"* or *"roll back the api deployment to the previous revision"*. Claude picks the right tool and calls it.

Prefer the terminal, scripts, or a coding agent in your editor? The [`okctl` CLI](/docs/cli) covers the same ground and ships an agent skill. The MCP server and the CLI are two front doors onto the same Ownkube API.

## What your agent can do

The MCP server mirrors the Ownkube API surface. Your agent can read and act across the resources you own.

**Inspect (read):**

| Area | What the agent can see |
| --- | --- |
| Organizations | The organizations your account belongs to, and your user profile |
| Clusters | Clusters, their status, and lifecycle state |
| Deployments | Deployments, status, revision history, and live activity |
| Environments | Environments and their variables |
| Registries | Connected container registries |
| Cloud accounts | Connected AWS accounts and their onboarding state |

**Act (write):**

| Area | What the agent can do |
| --- | --- |
| Clusters | Create, cancel, and destroy clusters |
| Deployments | Create and delete deployments, update config, upgrade platform versions, roll back, and promote one deployment's config onto another |
| Environments | Create and update environments and set their variables |
| Jobs | Trigger, cancel, and list job runs |
| Cloud accounts | Connect and reconnect AWS accounts |

Write actions change real infrastructure. Tools that create, upgrade, roll back, or destroy are described so the agent confirms with you before running them, but you are in control: treat an agent with this connector the way you would treat a teammate with dashboard access.

Two things the agent **cannot** do: it never returns live database connection credentials, and it never resets a database password. Those operations hand back a raw secret, so they stay in the [`okctl` CLI](/docs/cli) where a human runs them directly. An agent has no tool for either.

## Security model

- **OAuth, not keys.** The connection is authorized with OAuth in your browser. There is no long-lived key to leak, and you can revoke the connector from Claude at any time.
- **Scoped to your account.** The agent sees exactly what your account can see, in the one organization you approved. The same tenant isolation that guards the dashboard guards every tool call.
- **Your cloud stays yours.** Ownkube never receives your AWS or GCP keys through this server, and the agent cannot read them out. Cloud access is the scoped, least-privilege role you granted during [AWS onboarding](/docs/guides/connect-aws).
- **Auditable.** Every action runs through the same API as the dashboard and CLI, so writes are recorded the same way.

## Multiple organizations

The connection is pinned to the organization you picked at approval time. To act on a different organization, add the connector a second time and choose the other organization when prompted, or remove and re-add it. Read-only questions about *which* organizations you belong to work regardless of the pinned one.

---

**Don't see a feature you need?** Email [support@ownkube.io](mailto:support@ownkube.io?subject=Feature%20request). Ownkube is shaped by the teams using it and we ship what our users ask for.