Step-by-Step: Connecting Claude to CloudClawer

Two real paths to wire up CloudClawer's MCP server — the Claude.ai custom connector for everyday chats and the npm CLI for terminal automation — plus the pitfalls each one tends to trip on.

What CloudClawer's MCP server actually exposes

CloudClawer runs an HTTP MCP server at https://api.cloudclawer.com/mcp and ships an equivalent local stdio server in the cloudclawer npm package. Once connected, Claude gets a handful of container-shaped tools: launch, wait, status, stop, bash, log, tools, tool_config/ tool_set, secrets_*, proxy_allowlist_*, instructions_*, and apikey_status. The full reference lives in the docs.

Prerequisites

• A CloudClawer account — sign up here.
• For the Claude.ai path: a Claude.ai account.
• For the CLI path: Node.js 18+ on your machine.

Path 1: Add CloudClawer as a Claude.ai custom connector

This is the recommended path for everyday use — every Claude.ai conversation can drive your container, no terminal needed.

1. Open Settings → Connect in CloudClawer. The page shows three values: Server name, MCP server URL, and Client ID. Each has a copy button. The Client ID differs between the dev and prod environments, so always copy it from this page.
2. Go to claude.ai → Settings → Connectors and click Add custom connector.
3. Paste in the three values you copied from CloudClawer, save, and approve the OAuth prompt that opens.
4. In a new conversation: "Use CloudClawer to launch a container and run uname -a in bash."

Path 2: Install the CLI

The npm package gives you a CLI, a stdio MCP server, and the Playwright recording helper. Install once, authenticate once:

npm install -g cloudclawer
cloudclawer init

cloudclawer init opens your browser, you approve the new key, and the CLI writes it to ~/.cloudclawer with mode 0600. From there, every CLI command (and the stdio MCP server) reads that file automatically.

Verify with three commands:

cloudclawer launch
cloudclawer wait
cloudclawer bash "uname -a && whoami"

Path 2a: Wire the stdio MCP into Claude Code

If you want Claude Code (the desktop / IDE CLI) to expose the same tools, point it at the bundled stdio server. Edit ~/.claude/settings.json:

{
  "mcpServers": {
    "cloudclawer": {
      "command": "node",
      "args": ["${HOME}/.npm-global/lib/node_modules/cloudclawer/mcp-server.js"]
    }
  }
}

The server reads CLOUDCLAWER_KEY from ~/.cloudclawer — no header config required. Restart Claude Code so it re-reads the file.

Verifying the connection

The cleanest check is to ask Claude to actually use a tool, not just list one. From Claude.ai or Claude Code:

"Use CloudClawer to check container status. If nothing is assigned,
launch and wait. Then run `uname -a` in bash and show me the output."

Claude should call status (and if needed launch + wait), then bash. The container log line will also appear on your routines page.

Common pitfalls

Next steps

Now that Claude can talk to your container, set up a process definitionto gate risky tool calls — that's where CloudClawer earns its keep over a raw shell.