Poorly designed tool catalogs confuse models and operators alike. The Model Context Protocol gives you two primitives — tools and resources — and using each correctly improves reliability.
Resources: Ground Truth Without Side Effects
Resources should expose data the model needs repeatedly: connected account IDs, currency settings, field definitions, or yesterday's cached performance summary. Fetching a resource never mutates vendor state.
Example: before analyzing campaigns, the model reads a resource describing available ad accounts and their time zones. That prevents off-by-one date errors in reports.
Tools: Verbs With Clear Scope
Each tool should do one job with explicit parameters:
Avoid mega-tools like manage_campaigns with opaque action enums. Smaller surfaces map cleanly to user intent and log entries.
Descriptions Matter
The model chooses tools from natural-language descriptions. Include constraints in the description: "Requires campaign_id from list_campaigns. Cannot exceed org budget safety limits."
MCP360 maintains 50–100+ tools per service with consistent naming prefixes (meta_, google_) so multi-platform prompts disambiguate correctly on Ad Platform MCP.
Read-Only Keys for Exploration
Issue read-only API keys while operators learn phrasing. Graduate to write tools after validating metrics match native ad UIs. See Claude Desktop setup for a safe first connection.