MCP Server for Meta Marketing API (Facebook/Instagram Ads). Provides access to advertising campaigns via the Meta Marketing API with tools for analysis and management.
The server defines 16 tools with complete Pydantic input models and visible JSON Schema constraints. However, there are significant gaps in consistency and production readiness: (1) Output schemas are not documented, tools return responses but the schema structure is inferred, not explicit. (2) Descriptions vary widely in quality; some are adequate but many lack clarity on WHEN to use the tool vs. alternatives or dependencies. (3) Parameter descriptions are mostly present but sometimes generic (e.g. 'Formato output' without explaining why format choice matters). (4) Error handling is not visible in the provided code; no evidence of recovery guidance, categorization, or validation error messages. (5) Security: API credentials are loaded via environment variables (good), but the code does not show rate limiting, audit trails, or permission gates. (6) Tool composition is reasonable, each tool has a single responsibility, but response fields are not validated to match downstream tool parameters.
Create a new ad set (ad group) within a campaign.
Create a new advertising campaign.
Generate detailed performance reports with demographic and geographic breakdown.
Get creative details (images, copy, headlines, descriptions) for a specific ad.
Get performance metrics and insights for accounts, campaigns, ad sets, or individual ads.
List all advertising accounts accessible with the provided API credentials.
List all ads for a given ad set.
Output schemas not documented. Tools return data but the response structure is not explicit in the schema. LLMs cannot plan downstream tool calls without knowing what fields to expect.
Descriptions lack WHEN/WHY guidance. 'List all ads for a given ad set' does not explain when to call this vs. get_insights, or what the return structure enables. LLMs struggle with tool selection without contrastive context.
Parameter descriptions are minimal. 'Formato output: markdown per lettura umana, json per elaborazione' does not explain why an agent should choose one format over another. LLMs will guess arbitrarily without guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 45 | 2026-07-28+ | v2 |
List all ad sets (ad groups) for a given campaign.
List all campaigns for a given advertising account.
Pause or resume a specific ad.
Pause or resume an ad set.
Pause or resume a campaign.
Update targeting parameters for an ad set (age, gender, locations, interests, behaviors).
Update the bid amount for an ad set or ad.
Update the daily budget for an ad set.
Validate targeting specifications before creating or updating an ad set.
Complex parameter objects lack field documentation. 'geo_locations' in update_adset_targeting is an object with unspecified keys (countries, regions, cities, postal_codes), LLMs cannot construct valid requests without explicit field schemas.
No error handling visible. Code does not show recovery guidance (e.g., 'If targeting validation fails, call validate_targeting() first'), categorization (retryable vs. user-fixable), or actionable error messages.
Destructive/write operations (pause_resume_*, update_*, create_*) have no confirmation or dry-run pattern. Agents could accidentally pause campaigns or reduce budgets without a safety gate.
Tool naming for pause/resume is ambiguous. 'pause_resume_campaign' does both, better as 'pause_campaign' and 'resume_campaign' so intent is unambiguous from the name alone. The 'status' parameter is then unnecessary.
Response field naming not validated against downstream tool expectations. If create_campaign returns 'campaign_id' but another tool expects 'id', the LLM must reason about field mapping, increasing error rates.
No pagination guidance visible. list_campaigns, list_adsets, list_ads have a 'limit' parameter (1 - 100) but no documentation of total count, next_cursor, or what happens when results exceed the limit. Large datasets will truncate silently.
No audit trail or permission checks visible in code. Tools that modify ads (update_*, pause_*, create_*) should log who did what and when, and verify agent has authority. Compliance and incident response require this.