Claude Code

claude mcp add --transport http marina https://mcp.marinahost.app/mcp

That is the whole setup. Start Claude Code, run /mcp, and sign in; Marina’s tools appear. There is no token to mint and nothing to install, and you can withdraw the access from Marina later.

Running it yourself instead

For an assistant confined to your own machine, or one holding a credential you minted:

claude mcp add --env MARINA_API_TOKEN=<token> marina -- marina mcp

Mint the token first, so Claude is not holding your own credential:

marina tokens create claude --preset agent --spend 200

Where it gets stored

claude mcp add writes to the current project unless you say otherwise. --scope user makes it available in every project on the machine:

claude mcp add --scope user --env MARINA_API_TOKEN=<token> marina -- marina mcp

Sharing it with a repo

--scope project writes a .mcp.json at the repo root, which you check in so everyone gets the same tools. Never put the token in that file. Claude Code expands environment variables, so reference one instead:

{
	"mcpServers": {
		"marina": {
			"type": "stdio",
			"command": "marina",
			"args": ["mcp"],
			"env": {
				"MARINA_API_TOKEN": "${MARINA_API_TOKEN}"
			}
		}
	}
}

Each person exports their own MARINA_API_TOKEN, and the file carries no secret. Claude Code asks for approval the first time it sees a project-scoped server.

Without a token

If you leave --env off, the server uses whatever marina login stored on that machine, which is your own account with all of your permissions. Fine for a scratch project, wrong for anything you leave running. Give it a credential.

When it does not connect

claude mcp list

A failed-to-connect status next to Marina almost always means marina is not on the PATH Claude Code was started with. Run which marina and use the absolute path in command if they disagree.