Let me rewrite this headline following the rules. The original is "How we built our FinOps MCP server: lessons learned"

Short summary

Costory shares lessons from building a FinOps MCP server, revealing that exposing internal APIs directly to Claude failed because tools matched database shape rather than how models reason. They rebuilt with a single composable query tool where the model combines primitives like cost, metrics, and formulas instead of calling rigid per-question endpoints. Key takeaways include simplifying filter formats, shipping tools not answers, and designing error semantics that guide the model toward correct usage.

• •Exposing internal APIs as MCP tools fails because LLMs need simple contracts, not database-shaped endpoints
• •Rebuilt around a single composable query tool where the model combines primitives (cost, metric, formula) instead of calling rigid per-question endpoints
• •Lessons on filter format simplification, error semantics, and context injection for better agent-tool interactions

Generated with AI, which can make mistakes.