MCP for Agent Commerce: What Could Go Wrong

Alexander Polyakov and the Tutu team released an MCP server to let AI agents search real travel data instead of parsing web pages. The server must distinguish between agent and browser requests: POST/SSE for agents, HTML landing page for humans. Tool descriptions should be self-contained with readOnlyHint, idempotentHint, and openWorldHint annotations. Aggressive context compression reduced typical token usage from 133K to 34K (74%) by paginating results and truncating verbose data. Legal requirements force the final purchase step to happen on the merchant site: the agent selects options but the user accepts the offer, enters personal data, and pays. Authorization supports both OAuth 2.1 (for cloud clients like ChatGPT/Claude) and static key in header (for local clients); keys must never be in URL query strings. Testing uses evals: running live agents on 60+ test scenarios with a second model evaluating behavior. Rate limiting applies only to tool calls, not to the landing page or static assets, to avoid blocking human visitors. The key lesson: MCP servers must serve both human and agent audiences, with clear documentation and self-explaining tools.