Skip to content

ad-network-mcp-write-safety

Ad-Network / API MCP Writes: Money-Risk Guardrail

Section titled “Ad-Network / API MCP Writes: Money-Risk Guardrail”

Auto-loaded for marketing work. A mutating call through an MCP server or external API that touches a live ad network, monetization platform, MMP, or paid service can move real money. Treat every such write as high-risk. Applies to AdMob, AppLovin MAX, IronSource/LevelPlay, Mintegral, Unity Ads, Meta Audience Network, ad MMPs, and any billing/ads/payments API.

  1. Default to READ-ONLY. Prefer list_* / get_* / *_report tools. Reach for a write ONLY because the user explicitly asked for that exact change.
  2. Never fire a mutating tool without specific, explicit user confirmation that names: the operation, the target (account / campaign / offer / placement), and the before → after values. A generic “go ahead” is not enough for a spend-affecting change.
  3. Confirm the account/environment first. Verify the credentials point to the intended account and that it’s production vs. a designated test account BEFORE any write.
  4. One change at a time. No bulk or looped mutations without explicit batch approval (see preview-first-batch.md).
  5. Respect guarded-write gates. If an MCP exposes a confirmation/guard step (the ads-mcp servers do), never bypass it.
  6. Uncertain about spend impact? STOP and ask. If you can’t state how a parameter affects budget/bid/spend, do not send it.
  7. Never blindly retry a failed write. A timeout/error may have partially applied; re-read state before retrying.
  • Pure read/report tools: proceed normally, no extra confirmation.
  • A sandbox/test account the user has explicitly designated as safe for writes.
  • Removing a policy-flagged creative that cannot spend, via t1k-marketing-ua-how-to-optimize: allowed without per-write confirmation ONLY when adWriteSafety.zeroSpendCreativeRemoval.enabled is true in t1k-config-marketing.json AND the target account’s customer_id is in its allowedAccounts — report the operation, target and before → after after executing. Either condition absent → ask first. A creative that is still spending is never covered by this exception.

Rationale, the classify/confirm/verify/execute procedure, and a worked example: skills/t1k-marketing-monetization-mintegral/references/ad-network-write-safety-procedure.md.

  • ai-driven-design.md § “Side-effects with safety constraints”
  • preview-first-batch.md: smoke-test + confirm before any bulk operation
  • development-principles.md § “Errors Over Silent Fallbacks”