How Bonnard Builds Agent-Friendly MCPs

Short summary

Bonnard's MCP tool design prioritizes agent readability over human presentation: discovery-first tools let agents learn schema before querying, a small set of narrow tools prevents attention-budget overload, and compact responses with completeness flags avoid context-window poisoning. Errors carry actionable fixes so agents self-correct, and deterministic formatting work is pushed into code rather than left to the model.

• •Discovery-first tools let agents read schema and examples before acting
• •Small, narrow tool sets outperform one-firehose or one-per-metric approaches
• •Compact responses with completeness flags and error-fix hints keep agents on track

Generated with AI, which can make mistakes.