Can Claude or ChatGPT run Facebook ads for you?
Short answer: Claude or ChatGPT can read ad data and perform supported changes when connected through an authorized connector. You should approve specific changes and retain budget and strategy decisions. OpenAI and Anthropic both stress human control over agents, and Meta's connector works within the permissions you grant; none of them presents a connector as a way to leave ad buying unattended.
Start with a task you can check in Ads Manager
Before delegating work, establish the answer you expect from Meta's native tools. In Ads Manager, choose one ad account, set a completed reporting period and inspect Campaigns, Ad sets or Ads at the appropriate level. Record spend, the result event and the account currency.
Meta's Ads Manager introduction describes this baseline for managing delivery and tracking performance. The assistant should help you investigate that work, with numbers you can reconcile and actions you can inspect.
For example, ask for a report on yesterday's purchase campaigns before asking which campaigns deserve more budget. If the assistant mixes lead and purchase results, uses a different date window or omits an account, resolve that first. A fluent explanation does not establish that the underlying comparison is valid.
What “run my ads” can mean
There are several distinct permissions hidden in that request:
- Read existing campaigns and explain their performance.
- Recommend actions without submitting them.
- Execute a named action after approval.
- Repeatedly monitor conditions and act without a new decision from you.
Decide which one you want before connecting accounts. An assistant that can read a report does not automatically have authority to change its campaigns. A tool that can submit an edit also does not establish a recurring monitoring service.
Meta's official Ads MCP overview includes reporting and ad creation and management tools. That establishes supported categories, not permission for an assistant to choose any strategy or spend any amount on your behalf.
For ChatGPT, use a connector and a plan documented for the actions you need (how to connect Meta ads to ChatGPT). Multitool's connector is documented for Claude only.
What the platform providers actually say
OpenAI's agent-building safety guidance recommends keeping MCP tool approvals enabled so users can review reads and writes. This is guidance for building agents. It should not be confused with a guarantee that every ChatGPT configuration asks before every action.
Anthropic's Trustworthy agents in practice describes risks from reduced human oversight and identifies keeping humans in control as a core principle. It explains how the model, tools, operating instructions and environment all affect an agent's behavior.
Meta's Ads MCP documentation describes access limited by the permissions you grant; it neither recommends handing over budgets without supervision nor forbids unattended buying. Our recommendation is simple: keep human decisions over live changes, with explicit scope and review.
What to delegate and what to keep
This is a proposed division of responsibility, not a universal list of connector capabilities. Confirm that your chosen connector supports each action.
| Task | Delegate | Keep a human decision |
|---|---|---|
| Reporting | Retrieve data and draft a summary | Confirm event definitions, dates and currency |
| Anomaly checks | Flag unusual spend, results or delivery | Decide whether the issue requires intervention |
| Pausing by a clear rule | Execute a previously approved, deterministic condition | Define threshold, scope and reactivation behavior |
| Comment triage | Group questions, spam and possible abuse | Approve moderation and sensitive replies |
| Strategy | Assemble evidence and alternatives | Choose the objective, offer and test design |
| New creatives | Draft concepts and review candidates | Approve product claims, rights and final assets |
| Budget scaling | Prepare a proposal with exact objects and amounts | Approve the spending decision |
| Policy questions | Find the relevant documentation and flag uncertainty | Review the offer and decide the next action |
Do not turn an anomaly flag into an automatic spending decision by accident. A buyer may pause a campaign because its tracking is broken, because its economics changed or because a test ended. Those reasons call for different restart conditions even when the initial pause looks identical.
A deterministic rule and an AI judgment do different work
A deterministic rule evaluates a specified condition and performs its configured action. Given the same inputs and settings, it follows the same logic. An AI judgment can compare context and suggest explanations, but the decision needs review when the task leaves room for interpretation.
Consider an illustrative stop-loss rule: pause eligible ad sets when today's spend exceeds 90 units in the account's currency AND purchase results equal zero. The threshold is an example, not a recommended budget. Define which accounts and ad sets qualify, and check what happens when later results arrive.
Multitool ADS rules run 24/7 at account, campaign, ad set or ad level. Conditions support AND/OR, today, yesterday or the last N complete days, country and name scope, and child-object conditions based on any, all or count. The documented action is pause, with Telegram notifications.
An object paused by a Multitool rule switches back on automatically within an hour if the condition no longer holds. That behavior is material to a stop-loss workflow. Review it before using the rule for something you intend to end permanently. Automated budget scaling is not a documented rule action.
Rules are set up in the product, not through Claude, and asking the assistant to watch an account does not create an always-running rule.
How approved changes work through Multitool's Claude connector
The Multitool Ads connector offers analysis across connected Meta accounts, creatives, non-delivering ads, comments and Business Manager access. You select workspace and permission scope when connecting.
For an ad change, Claude prepares a preview that sends nothing to Meta. The action runs only after your approval, and the result is returned for each object. Supported edits include status, budget, name, creative, targeting, bid strategy, start time and special ad category. These are changes to existing advertising; Multitool ADS does not create new ads.
An approval should identify the account, object ID, proposed change and relevant amount or status. After execution, inspect the result. If Meta rejects a change, the connector does not retry it automatically. Resolve the stated cause before approving another attempt.
For analysis, answers normally use data already synced to Multitool and state its freshness. Ask for a refresh when the period needs updating. Claude never receives your Meta passwords or access tokens.
Give an assistant a bounded first assignment
Try: “For yesterday, list the campaigns in these accounts with spend and purchase results. Keep currencies separate. Show the account and campaign IDs, identify missing data, and make no changes.”
After reconciliation, ask for a proposed action with current and proposed values. Approve that specific proposal, then review the returned outcomes. Keep a named buyer responsible for the decision and for any other rules affecting the same objects. Neither a connector nor a convincing explanation guarantees advertising results.
You can inspect the account and creative workspace behind this workflow in the Multitool ADS live demo using demo data, with no account needed.
Sources: Meta Ads Manager introduction and official Ads MCP overview; OpenAI agent safety guidance; Anthropic human-control principles; Multitool Ads connector for Claude.