MCP Routing: When Agents Build Their Own Infrastructure
Agents are using MCP servers for routing decisions, not just data retrieval. MCP servers are becoming infrastructure that agents compose on their own.
An MCP server developer noticed something unexpected: an agent was calling their server every 30 minutes to make routing decisions. Not to retrieve data. To decide which provider to route the actual data request to. MCP servers are becoming infrastructure that agents compose on their own, not just tools that developers wire up.
From tool to infrastructure
The original vision for MCP was straightforward: expose tools to agents so they can call APIs. But agents are using MCP servers for something more fundamental. They query status and pricing data from MCP servers to make cost-aware routing decisions in real-time. The MCP server becomes a routing layer, not just a data source.
How routing via MCP works
An agent connects to multiple MCP servers via tools/list. It discovers available tools, their descriptions, and their parameter schemas. When a user query arrives, the agent does not call a hardcoded tool. It evaluates which tool best matches the query context, checks if the tool's MCP server is healthy, and routes accordingly. If one server is slow, the agent transparently routes to another.
// Agent discovers tools from Scavio MCP:
// google_search, reddit_search, youtube_search,
// amazon_search, walmart_search, + 6 more
//
// For "best wireless earbuds reviews":
// -> agent picks reddit_search (community opinions)
//
// For "AirPods Pro 2 price":
// -> agent picks amazon_search (product pricing)
//
// For "how to set up noise cancelling":
// -> agent picks youtube_search (tutorials)
//
// No if-else routing code. The LLM decides.The composability pattern
Agents are chaining MCP servers together for multi-step decisions. Check availability on one MCP server, then use a different MCP server for the actual API call based on what is available. This is emergent behavior: no developer designed the routing chain. The agent discovered it could optimize its own workflow by treating MCP servers as composable infrastructure.
What this means for MCP server builders
If you build an MCP server, your consumers are not just developers configuring agents. Your consumers are the agents themselves. Agents will discover your tools, evaluate their fitness for each query, and route around you if you are slow or down. Tool descriptions matter more than documentation because agents read descriptions to make routing decisions. Uptime matters more than features because agents will deprioritize unreliable servers.
The implication
MCP servers are becoming the API gateway layer for AI agents. Scavio's MCP server at mcp.scavio.dev/mcp exposes 11 tools covering 5 search platforms. An agent does not need to know that Google and Reddit are different services. It just sees 11 tools with descriptions and picks the best one per query. The abstraction layer is the agent's reasoning, not a developer's routing logic.