The videos are not wrong. Connect Meta's Ads MCP to Claude, and you can ask it to pull a 180-day report, find the three ads that carry the account, and draft variants of them, in one conversation. That is real and it is worth setting up. What the videos skip is the part that matters once money is moving: what the connection is at the API level, and therefore what it cannot do for you at three in the morning.
Meta's Ads MCP really does pull reports and draft creative in a single conversation. What it does not do is watch the account while you are asleep. Four things are worth checking before it touches spend: which token it is acting as, enhancements that switch themselves back on, rate limits that look like bans, and whether it reads before it writes.
What the MCP actually is
It is the Marketing API with a token attached. Every read and every write Claude makes goes to Meta as that token, with that token's permissions, and it happens when Claude is running a turn in your chat. Three consequences follow from that, and none of them are hidden — they are just easy to forget when a demo looks like autopilot:
- Nothing runs between sessions. Close the chat and the account is back to whatever Meta's own automation does. Unless you or a tool schedule it, there is no loop.
- The token is the operator. If you connected with your personal login, Claude has everything your login has, across every account in your Business Manager.
- Meta's own automation keeps acting on the same objects. Advantage+ enhancements, automated rules and the learning phase do not know a chat session exists.
Four things to check before you trust it with spend
These are the questions people are asking under those videos. The answers carry the documentation field names so you can check them yourself.
1. Which token
Meta's setup path is a login, not a token: the connection sends you through Facebook Login for Business and asks for ads_mcp_management, ads_management, ads_read and business_management. Which login answers that prompt is the whole decision. The narrower option is a System User in Business Manager — quickest to create in the Business Manager tool, or via POST /act_{ad_account_id}/assigned_users — scoped to one ad account with one task: ADVERTISE writes, ANALYZE only reads, never MANAGE. Test whether the MCP takes that token on one account before you plan an agency around it. That token can only touch what you assigned, and every change lands in the account's activity history under that user's name. It records what changed and who changed it — not why, not which prompt. It is a receipt, not a brake, and your client can read it too. One client, one system user. If you manage several accounts, separate tokens is only half the answer — one connection carries one token, so several clients means several connections and a real chance of the agent reaching for the wrong one. The other half is on Meta's side: the consent screen grants read or manage, and Settings → Business Integrations is where a client revokes you. Check what yours are set to before you assume a scope.
2. Enhancements that switch themselves on
Advantage+ creative enhancements are a per-creative setting: degrees_of_freedom_spec → creative_features_spec, with one enroll_status per feature — text_optimizations, add_text_overlay, image_touchups, image_uncrop, video_auto_crop, enhance_cta, adapt_to_placement, description_automation, image_background_gen and more. If you do not want Meta rewriting your copy or generating headlines, set each one to OPT_OUT when the creative is created. Then GET the creative back and read creative_features_spec. The read-back is the only proof it stuck: the toggle in Ads Manager writes the same field, so a UI that looks right is not evidence the API agrees. So treat the read-back as a habit, not a setting: check it when you build the creative, and again the next time you touch it. Nothing checks it for you — not Ads Manager, not us.
3. Rate limits are not bans
A common worry is that connecting an agent gets the account restricted. Using the API is the sanctioned way in. What an agent does hit is throttling: error code 17 when the user's token exceeds its limit, code 80004 for Ads Management calls under business-use-case limits. Per Meta's documentation, throttled calls simply fail, and the X-Business-Use-Case-Usage header tells you when you can resume. The budget is smaller than people assume, and it does not scale with your spend — it scales with your ad count: on Standard Access, 300 + 40 × active ads per hour. Ad policy is a different system entirely, and it judges the ad's text, images, targeting and landing page, not who submitted them.
4. Read before write
Ask it to GET the object and show you the current values before any POST that changes budget, status or targeting. It sounds pedantic. It is the difference between “I raised the budget to $50” and “I raised the budget to $50 on an ad set that was in learning, at 2am, on a creative Meta had already rewritten.” A model working from a stale summary will happily do the second one — and so will a model that read an instruction hidden in the data it just pulled. Meta's own MCP documentation says it plainly: “Agents act on your behalf. Anything they can call with your granted scopes, they can call — including in response to malicious instructions hidden in tool outputs (prompt injection).”
The hours nobody is watching
Here is the limit that no prompt fixes. Meta's learning phase, budget pacing and creative fatigue run on Meta's clock, twenty-four hours a day. A chat session gives you good judgement for the minutes you are in it and nothing for the rest. For the rest, you have two options that already exist: Meta's automated rules, which are free and which Meta checks at least once every thirty minutes, and something that runs on a schedule — a cron job, a script, or a product. We sell one of those and it is at the bottom of this page; the point here is only that the gap is real and a chat session does not fill it.
If you have one account and you are happy to open the chat every morning, Claude plus Meta's own rules is a good setup. You may not need anything else. Rules have one edge worth knowing before you build them: “adjust budget” raises as readily as it lowers, from the same trigger. Make your budget-increase rules notification-only.
Where we sit
Everix is the “process that keeps running” version of the loop above, and it is deliberately narrower than a chat. Cuts and kills execute on their own with an audit entry; every increase in spend, and the first dollar of any new campaign, stops at a human. That asymmetry is the whole design. It does not make creative — you upload it. It is Meta only. And it will not have the conversation with you that Claude will; it is built to be quiet.
The honest summary
Set the MCP up. Use a system user, opt out of the enhancements you do not want and read them back, expect throttling rather than bans, and make it read before it writes. Then decide who watches the account when the chat is closed — Meta's rules, your own process, or a tool built for that gap. The mistake is not using Claude on your ads. It is assuming a session is a service.