The server provides 15 tools across Facebook, Google, TikTok, and Reddit advertising/content platforms. Tool names follow verb_noun patterns (search_ads, get_details) and are generally clear. Descriptions are well-written (average ~150 chars) and explain WHEN to use each tool. Most parameters have type constraints (enums for sort/status/media_type) and descriptions. However, output schemas are entirely absent from visible code, no documentation of what these tools return to the LLM. This is a critical gap for agent composition. Parameter descriptions are solid but could include format constraints (e.g., date formats, ID patterns). Error handling is undocumented. Some tools have optional parameters that could benefit from clearer mutual exclusivity guidance (e.g., facebook_ad_details accepts id OR url).

• Document output schema for each tool. At minimum, list the key fields (e.g., 'Returns object with fields: ad_id (string), spend (number), impressions (number), demographics (array), creative_text (string)). Specify array limits (e.g., 'max 50 results per page').
• Add error handling guidance to each tool description. E.g., 'Returns error if ad not found (404), suggest facebook_search_ads with keywords instead. Returns rate limit error (429) if >100 calls/min, agents should back off exponentially.'
• Explicitly state mutual exclusivity in tool descriptions. E.g., 'Provide EITHER id OR url, not both. If you have only a name, call facebook_search_companies() first to get pageId.'
• Add regex patterns and format strings to schema definitions. E.g., for country codes: pattern: '^[A-Z]{2}$' (ISO 3166-1 alpha-2). For dates: format: 'date' (YYYY-MM-DD).
• Clarify cursor pagination. E.g., 'cursor is an opaque token returned in paginated responses. Pass it unchanged to get the next page. Null cursor in response means end of results. Cursors expire after 24 hours.'
• Add result limits to descriptions. E.g., 'Returns up to 20 results per page (default limit=20, max=100). Use cursor to paginate large result sets.'
• Consider adding tool annotations (readOnlyHint=true for all current tools, since all are searches/gets). This signals to agents that these are safe to retry.