Your AI agent can make Facebook ads. Should it be allowed to spend on them?

Every reliable ad-running agent ends up with the same architecture: generate wide, judge independently, launch narrow.

• ✓Generate — the agent drafts 10–20 variants (its actual strength).
• ✓Gate — each variant goes through assess_ad_creative. Weak hooks, unclear angles, and likely policy rejections get fix_first or kill— deterministically, from server-side gates the model can’t talk its way past.
• ✓Launch — only run verdicts reach the Meta ads account. The agent fixes or regenerates the rest before any budget moves.

The judgment is platform-specific: the same copy is scored as a Meta feed ad, not as generic marketing text — hook strength in the first line, feed-native tone, Meta policy exposure.

The fastest path is the agent Skill — it works in Claude Code, Cursor, Codex, and Gemini:

npx skills add spendict/skills

Or connect over MCP (Claude, Cursor, any MCP client):

https://www.spendict.com/api/mcp

Building with a framework? There are copy-paste recipes for CrewAI, LangGraph, the Claude Agent SDK, and n8n in the integrations repo, and everything is mirrored over plain REST for custom stacks.

{
  "overall_score": 52,
  "launch_recommendation": "fix_first",
  "dimension_scores": {
    "hook": 38, "angle": 71, "clarity": 74,
    "audience_resonance": 66, "platform_fit": 62,
    "cta": 80, "compliance_safety": 90
  },
  "predicted_failure_mode": "Hook reads like a product announcement — scroll-past risk.",
  "prioritized_fixes": [
    { "issue": "Hook below gate (38)",
      "suggestion": "Lead with the audience's cost of inaction, not the product name." }
  ]
}

Structured, machine-readable, and consistent — the same ad always gets the same verdict, which is what makes it safe for an agent to automate against. Beyond creative, the agent can also call audit_campaign_structure before building the campaign, strategize_targeting to design the audience plan, and analyze_campaign_performance once real metrics exist.