anthropics / anthropics/claude-cookbooks
Cookbook proposal: marketing analyst agent on Meta's official MCP with a numeric verification grader
- Dominant language
- Jupyter Notebook
- Stars
- 52.7k
- Forks
- 6.3k
- Avg merge
- 25m
- Merged PRs (30d)
- 6
Description
Meta's Ads AI Connectors shipped in open beta on April 29, 2026. Their official MCP server uses standard Business OAuth with no developer app or app review, supports Claude, and exposes ~29 tools including `get_campaigns`, `get_insights`, and `update_ad_set`. Reproducibility is now straightforward.
I'd like to contribute a cookbook in the Managed Agents category that fills two gaps I see in the current catalog:
1. **No cookbook shows a Claude agent using a vendor MCP server end to end.** All current Managed Agents examples are offline or use Anthropic-provided tools. Vendor-MCP integration is one of the most-asked-about production shapes and Google, Meta, and several others have shipped MCP servers in the last six months.
2. **No cookbook on numeric verification.** Cookbook #2 (*Outcomes: agents that verify their own work*) covers quotes and citations. Marketing reports fail most often when a number is wrong. A stateless grader that re-derives every figure from the raw rows before the report ships is, as far as I can tell, missing.
**Proposed shape** (~280 lines, 7 sections):
- Read-only writer agent on `claude-sonnet-4-6` calls Meta's MCP for current and prior period insights, writes a comparison HTML report with plotly charts.
- Stateless grader on `claude-haiku-4-5` loads the report and the source rows, re-derives every numeric claim, returns pass and a list of unverified claims on fail.
- Writer revises on failure, loop ends on pass or after N rounds.
- Read-only tool scope (no `update_ad_set` or other mutating Meta tools), explicit in the system prompt and `permission_policy`.
- Fixture-first: ships with a captured Meta `/insights` response so the notebook runs in CI without an ad account. Live mode documented as an appendix.
Differentiation from cookbook #5 (*data analyst agent*): live API via vendor MCP rather than a mounted CSV, period-over-period comparison rather than single-window analysis, and the numeric verification loop on top.
Two questions before I invest the time:
1. Is this overlapping with anything in flight internally? Given the launch date, I'd rather check before writing.
2. Is the read-only scope and fixture-first design the right way to thread the safety and reproducibility constraints, or would you prefer a different framing?
Happy to adjust the angle if there's a better gap to fill.
Contributor guide
Assessment
This issue has not been assessed yet.