AI agents & MCP

Smallcloud treats agents as first-class deployers. The bundled MCP server exposes the full app lifecycle as tools over stdio.

How do I connect Claude Code to Smallcloud?

Sessions started in the Smallcloud repo pick it up automatically from .mcp.json. For any other project, add:

{
  "mcpServers": {
    "smallcloud": {
      "command": "node",
      "args": ["/path/to/smallcloud/packages/mcp-server/dist/index.js"]
    }
  }
}

What tools does the agent get?

• deploy(sourceDir, appName, ownerEmail?) → the public URL plus a single-use sign-in link.
• list() → every app with status and URL.
• logs(appName, tail?) → recent app output.
• delete(appName) → stops the app, removes its URL and data.

What does the flow look like in practice?

You say: "build a link-shortener and put it on smallcloud." The agent writes the app, calls deploy, reads logs if something is off, fixes and redeploys, then hands you the URL and sign-in link. The app comes up inside the same sandbox as any human deploy: private, no egress, resource-capped.

Why is this safer than giving an agent a cloud account?

The blast radius is structural. The MCP tools can only deploy/list/log/delete sandboxed apps on your box — there is no billing to run up, no IAM to misconfigure, no public bucket to leak. Every action lands in the audit trail with its actor.