Seven rules for MCP tools that an AI assistant cannot misread

Description and use

Seven rules for designing an MCP server to prevent language model misinterpretation of health data APIs.

Problem

Language models misinterpret API tool lists, leading to incorrect or misleading answers when interacting with health data APIs.

Outcome

The MCP server now reliably presents accurate, non‑misleading information to language models, preventing erroneous dosage calculations, misinterpretation of empty fields, and accidental data writes.

Workflow

• 1. Reduce tool count: merge similar endpoints into single tools per question type.
• 2. Document empty field meanings in tool descriptions.
• 3. Explicitly forbid unit conversions and arithmetic in descriptions.
• 4. Label numeric fields with source (studied vs. printed).
• 5. Ensure all tools are read‑only and state this in the description.
• 6. Scope tools by OAuth scopes (evidence.read vs. me.read).
• 7. Return concise summaries and external links, not full licensed text.

Evidence

When has_confirmed_label is false there is no panel, and that means the match could not be confirmed. It does NOT mean the product has no ingredients. Never tell a reader a product contains nothing.

Source

Our descriptions say: each amount keeps its own unit; do not add amounts across units or items; do not convert IU into a mass.

Source

The study tool returns the title, year, participant count, our summary and the PubMed link.

Source