Problem

Currently, MCP servers configured under mcp.servers in openclaw.json are only accessible to embedded Pi / coding agent runtimes, not to the main OpenClaw agent.

The openclaw mcp list command only manages the MCP server registry, not the actual tool exposure to the main agent. MCP server definitions stored via mcp.servers are consumed by embedded runtime adapters (like Pi), while the main agent communicates with models via model providers directly.

This means:

• Main agent image tool → model provider (e.g. Minimax API) → limited by model capability
• Main agent cannot use MCP tools like web_search, understand_image from MCP servers

Expected Behavior

MCP tools should be usable by the main OpenClaw agent in the same way they are usable by embedded Pi/coding sessions.

Proposed Solutions

Option A: Allow mcp.servers definitions to be used by the main agent tool dispatch layer
Option B: Add a config flag like mcp.exposeToMainAgent: true
Option C: Document the limitation clearly and provide best-practice guidance

Workaround (Current)

Use shell exec to call mmx CLI directly for image understanding and web search. This works but bypasses the MCP abstraction layer.

Submitted via OpenClaw agent.