MCP as the bridge to private APIs
Model Context Protocol is useful when it turns internal systems into safe, typed tools for agents — not when it becomes another abstraction layer.
Agent demos love public tools. Real products need private ones: inventory, pricing, booking, permissions.
That is where Model Context Protocol earns its keep. An MCP server can expose live B2B APIs to a local or remote LLM with explicit schemas, auth boundaries, and auditability — without stuffing secrets into prompts or reinventing tool-calling for every model.
On Travel Professor, the interesting work was not “call an LLM.” It was:
- Mapping Crux aggregator endpoints into stable tool contracts
- Keeping high-concurrency travel infrastructure behind a narrow, reviewable interface
- Letting the agent reason over multi-step hotel discovery without becoming a second API gateway
MCP is not magic. Bad tools still produce bad agents. But a thin, well-typed bridge between models and private systems is a better default than ad-hoc function calling scattered across the codebase.
Build the harness. Keep the surface small. Measure whether the agent still does the right thing when the catalog shifts.