MCP in Claude Code, Codex and Cursor

If your AI works in a folder on your computer, as Claude Code, Codex and Cursor do, it can build that folder and put it online itself. It does so with the local MCP server in the Cubli CLI. On top of that, it can do everything the connector for Claude and ChatGPT can: projects, databases, domains, logs and rolling back.

Claude Code

One line in your terminal:

claude mcp add cubli -- npx -y @cubli/cli mcp

Codex, Cursor and other tools

Add an MCP server that runs locally (stdio), with npx as the command and -y @cubli/cli mcp as the arguments. In Cursor, put this in .cursor/mcp.json in your project, or in ~/.cursor/mcp.json for all your projects:

{
  "mcpServers": {
    "cubli": {
      "command": "npx",
      "args": ["-y", "@cubli/cli", "mcp"]
    }
  }
}

You need Node.js 20 or newer.

Log in

On first use the tool reports that you are not logged in, and calls login itself. Your browser opens: check the code, choose the workspace and click Allow access. Your AI then works only in that workspace.

Already logged in with npx @cubli/cli login? The MCP server uses that login. They share it.

What you can ask

“Put this folder online on Cubli, in project Bakkerij Jansen.”
“The deploy fails; read the build log, find out why and fix it.”
“Create a Postgres database and connect it to the app.”
“Set up staging, deploy there, and promote it to production when I agree.”
“Connect bakkerijjansen.nl and tell me which records to set.”
“Go back to the previous version.”

What only works locally

The local server has a few tools that need a folder on your computer:

Tool	What it does
`deploy`	Packs a folder (respecting your `.gitignore`, never `.env` or `node_modules`), builds it without a Dockerfile and waits until it is live. The first time it needs a project; after that the folder is linked. Can deploy to staging too.
`check_app`	Reads a folder before its first deploy: which framework, what the app uses and which variables it needs. Nothing is uploaded.
`pull_code`	Gets the live code into a folder, for example to change an app made from a template.
`login`	Logs in through your browser, or switches workspace.

If a deploy fails, your AI gets the reason and the last lines of the build log, so it can fix it. The previous version stays live meanwhile.

Rather not install anything?

Claude Code and other tools can also use the connector at https://mcp.cubli.io, over HTTP. Your AI can then do everything except deploy a folder:

claude mcp add --transport http cubli https://mcp.cubli.io

Then type /mcp in Claude Code to sign in. In Cursor, VS Code and other tools, add an MCP server of type HTTP (or “streamable HTTP”) with that address.

What your AI may and may not do is in What an AI tool may do.

Was this helpful?