MCP server
One URL gives an agent app its own desktop. No code, nothing to install.
Kubi runs a hosted MCP server with the API. Add one URL to Claude Code, claude.ai, Claude Desktop, Codex, Cursor or VS Code, sign in, and the agent can see a screen, click and type on it, run commands and move files:
Code
It speaks MCP's Streamable HTTP transport. It runs inside our API, so there is nothing to install or keep running on your side.
MCP, the SDK or the API
All three reach the same screens, with the same rules and the same history.
| You have | Use |
|---|---|
| An agent app: Claude Code, claude.ai, Codex, Cursor, VS Code | The MCP server |
| Your own code in TypeScript | The TypeScript SDK |
| Your own code in any other language | The API |
How a session goes
- The agent calls
screenshotand gets the screen as an image with itsscreenshot_id. - It calls
actwith up to 20 actions and thatscreenshot_id. Actions go by point, or by an element id fromread_tree. - The reply carries the screen after the actions, with a new
screenshot_idfor the next step. One call is one step. - If anything ran on the screen since the screenshot, a person included, the actions don't run and the agent gets
STALE_OBSERVATION: it looks again. The full loop: The look and act loop.
The server's instructions are a short guide to that loop, and guide answers more by topic. read_tree with window: "all" finds an element in any window, apps lists apps for act's launch, page reads and fills a browser tab by ref, and run_script runs a GUI script inside the screen for loops in one call. Besides that loop, exec runs shell commands, read_file and write_file move files, clipboard reads or sets the clipboard, and viewer_link gives a person a live view of the screen.
What it never does
The MCP server has no tools to create or delete screens, make keys or touch billing. Those stay in the dashboard, for people. Starting, stopping and restarting a screen aren't offered yet either; do them in the dashboard or through the API.
In this section
- Connect: the setup for each agent app, or an API key instead.
- Sign-in and permissions: what you grant, for how long, and how to take it back.
- Tools: every tool, with its inputs, read from the server itself.
- Errors and limits: what a failed call says, and the sizes and times to plan for.