mcpmux implements a two-stage tool discovery gateway with 5 tools. Tool naming follows verb_noun convention (discover_tools, call_discovered_tool, refresh_tool_catalog, gateway_stats, upstream_status). Descriptions are present but brief (30-80 chars), below the 50-200 char LLM-optimized baseline. Input schemas are visible for discover_tools and call_discovered_tool with proper type definitions and descriptions. However, refresh_tool_catalog, gateway_stats, and upstream_status have empty input objects with no documented output schemas. No error handling guidance, no pagination support, and no output schema documentation across any tool. Tool composition is reasonable (discovery → execution pattern), but lacks chaining IDs and recovery guidance.

• Expand tool descriptions to 100-150 characters. For discover_tools: 'Search upstream MCP servers for tools matching a natural language query. Returns tool IDs and brief descriptions without full schemas. Call this first to narrow the tool space before calling call_discovered_tool.'
• Document output schemas for all tools. E.g., gateway_stats should return {successCount: number, failureCount: number, topTools: [{toolId, successCount, failureCount}]}. upstream_status should return {servers: [{id, connected, lastError}]}.
• Add error handling guidance to discover_tools and call_discovered_tool. E.g., 'If no tools match the query, try a broader search term or call refresh_tool_catalog to reload the upstream inventory.'
• Clarify the semantics of allowHighRisk and includeDomains in discover_tools. Document what domains are available and what 'high-risk' classification means (e.g., destructive, requires auth, external API).
• Add a tool to retrieve detailed schema for a discovered tool (e.g., get_tool_schema(toolId)) so agents can inspect parameters before calling call_discovered_tool.