Kubme plugs into the AI assistants you already use (Claude, ChatGPT, Le Chat, Gemini, Copilot, Cursor…) through the MCP protocol (Model Context Protocol). Once connected, your assistant reads and edits your Kubme components, just like Kubme AI.

From your usual assistant, ask for example:
The assistant acts on your behalf, with your rights: it only sees and edits what you can see and edit. While it works on a component, it shows up among the people present, for example “Claude, for Julie” (see Co-editing).
The first time, your assistant opens a Kubme sign-in window: sign in, then accept the “Kubme AI assistant” access. You only do this once.
The MCP server works with any model that can use tools, in any client that accepts a remote MCP server with OAuth. Prefer your client's most capable model: small models are worse at chaining tools.
If you belong to a single organization, it is picked automatically. Otherwise your assistant asks which one to work in. You can also tell it: “Work in the HDM-SOFT organization.”
Sensitive actions (delete, share, publish) need your approval, just like in Kubme AI: see Approvals. Your assistant shows the request, and you approve or decline it.
If your client cannot show these requests (MCP “elicitation”), sensitive actions are refused with a message asking you to do them in the studio. Everything else works normally.
Each tool used by an external agent counts as one MCP action. The number of actions is limited per day and per organization, depending on your plan.
When the limit is reached, the assistant tells you: nothing is done until the next day, unless you move to a higher plan.
Kubme can temporarily pause AI agent access. The assistant then answers “Kubme is not available to AI agents for now”: try again later.
The full tool list is in MCP tools.