Smart Router Overview

Smart Router exposes one mcp tool that routes across all configured servers. Instead of registering every backend tool directly, the model calls mcp(action=..., server=..., tool=...) and the bridge resolves execution.

For multi-server setups this dramatically reduces tool-definition tokens. The README positions router mode at roughly ~98-99% token savings versus large direct-tool surfaces.

At a glance

• Intent routing from natural language requests
• Schema compression to reduce token footprint
• Result caching with LRU + TTL
• Batch calls for parallel execution
• Multi-server resolution when tool names collide
• Adaptive promotion for frequently used tools

Text architecture diagram

MCP client (Claude/Cursor/OpenClaw)
            |
            v
      single "mcp" tool
            |
            v
  Smart Router (intent + resolution + batching + cache)
            |
            v
    Security layer (trust/filter/limits)
            |
            v
  Upstream MCP servers (todoist/github/notion/...)

With vs without smart router

Setup	Model-visible tools	Context impact
Without smart router (direct mode)	Every backend tool is exposed	Grows quickly with server count
With smart router (router mode)	Single mcp meta-tool	Much smaller tool-definition footprint

Base router call patterns

mcp(action="list", server="todoist")
mcp(action="call", server="todoist", tool="find-tasks", params={"query":"today"})
mcp(action="intent", intent="find my tasks for today")

Jump to feature pages in this section for behavior and config knobs.