Skip to main content

Looker Code Mode MCP Server

lkr code-mode allows you to invoke a Python-based Model Context Protocol (MCP) server. It offers an AI agent the unique capacity to batched-execute Python commands securely within the Monty sandbox against your active Looker instance!

How It Works​

Instead of dumping hundreds of explicit tool declarations onto your AI agent (token bloat), lkr code-mode exposes exactly one tool: run_python_code(code: str).

The tool instantiates Looker SDK natively, searches all bound methods, and passes them safely onto Monty's environment as global functions. When the LLM writes standard Python code (e.g., me(), folder(id)), Monty will process it correctly locally!

Key Features:​

  • 100% Tool Coverage: Accesses all Looker SDK public operations smoothly without token limits.
  • Recursive Translation: Complex Looker models like User, Folder, Dashboard get string-converted into dictionaries immediately before ingesting them into Monty.
  • Automatic PKCE Restarter: Caught an invalid token? Code Mode immediately catches InvalidRefreshTokenError and safely opens up your PKCE authentication browser automatically.
  • Extremely Secure: Monty interpreter ensures isolated sandbox processing. No local filesystem accesses are exposed.

Continuous Usage​

1. Direct CLI Execution (Sandbox Mode)​

If you want to execute Python scripts or inline code directly against your Looker instance without setting up an MCP client, use the sandbox command:

# Execute inline Python code
uvx lkr-dev-cli code-mode sandbox --code="return me()"

# Execute Python script file
uvx lkr-dev-cli code-mode sandbox --file=./path/to/script.py

2. Starting the Server​

To immediately trigger the MCP stdio listener, use:

uvx -q lkr-dev-cli[codemode] code-mode run

2. Client Configuration​

To hook this server into Cursor or Claude Desktop natively over stdio, append the following onto your mcpServers configuration JSON. Make sure to pass your Looker instance credentials as environment variables (see standard API requirements in README.md):

{
  "mcpServers": {
    "lkr_codemode": {
      "command": "uvx",
      "args": ["-q", "lkr-dev-cli[codemode]", "code-mode", "run"],
      "env": {
        "LOOKERSDK_BASE_URL": "https://your.looker.instance",
        "LOOKERSDK_CLIENT_ID": "your-client-id",
        "LOOKERSDK_CLIENT_SECRET": "your-client-secret"
      }
    }
  }
}

3. Visual Inspector​

To check things out on a web panel:

npx @modelcontextprotocol/inspector uvx -q lkr-dev-cli[codemode] code-mode run

Architectural & Troubleshooting Notes​

The stdio Stream Corruption Trap​

When running an MCP server over standard input/output (stdio) transport, JSON-RPC packets are transmitted directly through the process's stdout.

If Python code executed within the server (such as tool execution or third-party libraries) writes raw text directly to sys.stdout via standard print() calls, that raw text corrupts the JSON-RPC response stream. This causes fatal JSON parsing errors on the client (e.g., invalid character 'C' looking for beginning of value) and immediately closes the connection.

To safeguard against this:

  1. run_python_code automatically wraps all Monty execution in a redirect_stdout block.
  2. If any print() output is captured, it is safely packaged and returned as part of a valid JSON structure rather than polluting raw stdout.