Written by Uptimal, the team that builds and operates Upbuild and runs marketing analytics for multinational clients across the globe. Everything below reflects how the connector actually behaves in production.
The fear is reasonable: “give the AI my ad accounts” sounds like handing over the keys. The reality, with a properly built MCP connector, is narrower and more auditable than most agency access you have already granted. Here is the model, layer by layer, and the questions that separate serious vendors from wrappers.
Layer one: what is authorized
Per-user OAuth against each platform, with read-only scopes you can verify on the consent screen itself: Meta shows ads_read and read_insights, Google shows read-only Ads and Analytics scopes, TikTok is gated by a reporting allowlist. If a vendor cannot show you the scope list before you click approve, stop there. Write-capable scopes are not “convenient extras”; they are a different risk class.
Layer two: where credentials live
Tokens are stored encrypted, per user, and die on disconnect: both by removing the connector and by revoking the app inside the platform (Facebook settings, Google account permissions), which invalidates tokens at the source. There is no shared service account whose blast radius spans customers, and no password ever transits: OAuth means the platform authenticates you, the connector only holds the resulting scoped token.
Layer three: what the model actually sees
The assistant sees tool responses: the report rows your question required, inside your conversation, governed by the assistant vendor’s data terms (both Anthropic and OpenAI offer no-training controls on the relevant plans). It does not see credentials, does not browse your account, and cannot exceed the token’s scopes no matter how it is prompted, the ceiling is cryptographic, not behavioral.
The grill-any-vendor checklist
- Show me the OAuth scopes per platform, pre-consent.
- Per-user tokens or shared service account?
- Encryption at rest, deletion on disconnect: stated where?
- Can any tool write? (The only acceptable answer for reporting products: no.)
- What appears in your logs, and for how long?
Upbuild’s answers live in the privacy policy and the evaluation guide; ask competitors the same five and watch the pauses.
Ask your ad platforms directly, in Claude or ChatGPT.
Upbuild is a read-only MCP connector for ten marketing platforms: 55 tools, live data, a 14-day free trial with no card.
Start free trial → Talk to usFrequently asked questions
Can a prompt injection make the assistant edit my campaigns?
Not through a read-only connector: the token lacks write scopes, so no instruction, malicious or otherwise, can exceed it. That guarantee is exactly why read-only should be non-negotiable for reporting tools.
Is my conversation data used to train models?
That is governed by your assistant plan: Anthropic and OpenAI both provide no-training controls on business tiers. The connector itself transmits tool responses into your chat and stores credentials, not conversations.