cc-openclaw-bridge: Instructions file for Claude Code

OpenClaw Bridge — Instructions for Claude Code

You have access to MCP tools from the openclaw-bridge server for communicating with the user when running in headless/non-interactive mode (claude -p).

When to use

Always use these tools instead of AskUserQuestion when running headless. AskUserQuestion does not work in -p mode — these tools do.

Tools

ask_user

Ask the user 1-4 questions. Supports rich options with descriptions.

ask_user(
  questions: [
    {
      question: "Which database should I use?",
      header: "Database",
      options: [
        { label: "PostgreSQL", description: "Reliable, great ecosystem" },
        { label: "SQLite", description: "Simple, embedded, no server" }
      ]
    }
  ],
  context: "setting up the backend"
)

Rules:

• Batch related questions into one call (max 4) — fewer interruptions
• Always include header (short tag, max 12 chars)
• Add description to options — helps user decide without context-switching
• Include context — user may have multiple CC agents running
• If no fixed options, omit options for free-text answer
• Set multiSelect: true when multiple choices are valid

notify_user

Send a progress update without waiting for a reply.

notify_user(
  message: "All tests passing, creating PR...",
  context: "fixing auth bug"
)

When to notify:

• Starting a major phase (tests, build, deploy)
• Hitting an unexpected blocker
• Completing the task

Important

• The user may respond with vague answers like "whatever" or "you pick" — the OpenClaw agent will interpret and choose for them
• Questions are delivered to the user's messaging app (Telegram, Signal, etc.) — keep them concise and self-contained
• Don't ask unnecessary questions — if you can make a reasonable default choice, do so
• A summary of your session is automatically sent when you finish