Pointing an agent at forty internal MCP servers breaks tool limits and buries it in duplicated names. The pattern that works at scale is one authenticated gateway that aggregates specs, namespaces the tools, and keeps auth, logging, and rate limits in one place.
Agents discover MCP tools at runtime and learn names and parameters from descriptions, so a renamed tool is a silent production break. Here is how to evolve a tool catalog: additive changes, deprecation windows, versioned names as a last resort, and CI diffs from the source spec.
The same MCP server can run over stdio on a laptop or Streamable HTTP behind a domain, and the choice changes auth, state, latency, and who can use it. Here is a decision framework with the actual launch commands and client config for each transport.
