Can ChatGPT or Claude Manage Your Ad Accounts?

Not officially — and not fully. As of late July 2026, neither Google nor Meta offers a product that lets ChatGPT or Claude freely manage a live ad account by chat. Google publishes exactly one official MCP server for the Google Ads API, and it is explicitly read-only: three tools, no write access, and documentation that says it “cannot modify bids, pause campaigns, or create new assets”. Meta publishes no official MCP server that our research could find at all — the only sanctioned automated path into a Meta ad account remains the Marketing API, which is gated behind Meta App Review for production use.

What people are actually using today is a large, fast-growing ecosystem of unofficial, developer-built MCP servers that wrap the real platform APIs with your own credentials. A live query to the GitHub Search API during our research returned 312 repositories matching “google-ads mcp” and 36 matching “facebook-ads mcp” — and exactly one of the Google-side results belongs to Google itself. With every other repo, capability and safety depend entirely on a third party’s code, and for Meta specifically, handing your access token to an unapproved third party can itself violate Meta’s Platform Terms.

This article maps what officially exists, what the community has built, the myths worth killing, and the honest gaps we could not confirm — with primary sources throughout. If what you actually want is the outcome (AI running your ads) without wiring an agent to your ad account yourself, that is the job of a purpose-built AI ad manager; we cover where that fits at the end.

What “connecting an AI to your ad account” actually means

The plumbing behind almost every “manage ads from ChatGPT/Claude” setup is the Model Context Protocol (MCP) — an open-source standard for connecting AI applications to external systems, described by the official MCP site as working “like a USB-C port for AI applications.” Anthropic announced MCP on November 25, 2024 as an open standard for secure two-way connections between data sources and AI-powered tools, and per the protocol’s own site it is supported across Claude, ChatGPT, VS Code, Cursor, and others — it is not exclusive to any one vendor.

One telling detail: OpenAI’s own MCP documentation says servers intended for ChatGPT deep research and company knowledge should implement two read-only tools — “search” and “fetch” — and gives no advertising platform as an example. Across every official source we could verify, the default posture leans read, not write.

What Google officially offers — and what it refuses to do

Google ships two distinct AI developer tools for Google Ads. They are easy to conflate, and they have materially different purposes and safety models.

First, the official Google Ads MCP server, which Google describes as “a standardized bridge to the Google Ads API” that lets AI agents analyze and retrieve campaign data using natural language. Its documentation states “Mode: Read-only (current release)” and it exposes exactly three tools: list_accessible_customers (accessible customer IDs and account names), search (GAQL requests for resource metrics, budgets, and status), and get_resource_metadata (metadata about Google Ads API resource types). Setting it up requires a 22-character developer token, a Google Cloud project ID, and OAuth credentials — either an OAuth2 client ID/secret pair or application default credentials. There is no write tool. It cannot pause a campaign for you no matter how nicely you ask.

Second, the separate Gemini-powered Google Ads API Developer Assistant, which Google explicitly describes as “not a chatbot” but “a mission control system for Gemini-powered advertising engineering.” It is aimed at developers writing code against the API, not advertisers managing a live account — and on the question of changes, its install guide is blunt: “For safety, mutating operations (create, update, delete) are generated but never executed.” Google’s own AI tooling drafts the change and hands it to a human. It does not push the button.

What Meta officially offers

Meta’s Marketing API is the sanctioned way to “automate advertising across Meta technologies,” covering ad creation, management, and performance analysis. But access starts on a “Limited access” tier that Meta’s authorization docs describe as “Heavily rate-limited per ad account. For development only. Not for production apps running for live advertisers.” Full access requires passing Meta App Review, plus a usage bar: at least 500 Marketing API calls in the last 15 days, with an error rate under 15% across the last 500 calls.

As for an official Meta MCP server: we could not find one. Meta’s Marketing API documentation makes no mention of MCP, and a search of Meta’s own GitHub organisations surfaced no Marketing API MCP server. That is a “none found as of July 2026,” not a proven absolute — but it means every “Meta Ads MCP server” you can install today is unofficial.

Six myths about managing ads from ChatGPT or Claude

Myth 1: Google or Meta now offer an official AI connector that fully manages your campaigns

Neither does. Google’s only official MCP server is read-only with three tools, per its own documentation. No official Meta MCP server appeared anywhere in our research — Meta’s Marketing API docs never mention MCP, and the only sanctioned automated path for Meta remains the App-Review-gated Marketing API.

Myth 2: Any “Google Ads MCP server” you find online is vetted by Google or Meta

Our GitHub Search API query returned 312 repositories matching “google-ads mcp” and 36 matching “facebook-ads mcp” at the time of checking — nearly all independent community projects (the top facebook-ads result, for example, is gomarble-ai/facebook-ads-mcp-server) that wrap the real APIs with your credentials. Exactly one repo in the Google set belongs to Google’s own organisation: googleads/google-ads-mcp, the read-only server.

Myth 3: Sharing your Meta access token with a third-party MCP server is harmless

Meta’s Platform Terms (Section 6.a.iv) state: “You must protect and not transfer, share, or solicit Meta user IDs, access tokens, or app secrets,” with only a narrow exception for a Service Provider who helps you build, run, or operate your app. Handing a token to an unapproved third-party MCP server can itself be a Platform Terms violation — independent of anything the AI does with it. What follows is a description of what the platform terms say, not legal advice; if a compliance decision turns on it, have a professional read the terms with your specific setup in front of them.

Myth 4: Since it is “just API calls,” an agent can auto-adjust bids with no human checkpoint

Google’s Required Minimum Functionality policy states that all API Clients providing Creation or Management functionality must satisfy requirements related to bid adjustments, including: “End users must be able to review and edit the bid adjustments before it is set by the Google Ads API Client.” Human review of bid changes is a compliance requirement, not optional UX. We wrote more about drawing that line in what to automate and what to gate.

Myth 5: Claude’s or ChatGPT’s directories already have a ready-made Google Ads or Meta Ads integration

Claude’s Connectors directory itself states that “Connectors in the directory are built and maintained by third-party developers using the Model Context Protocol (MCP)” — they are not Anthropic-built. On page 1 of the 17-page directory (24 connectors, “10x Genomics Cloud” through “AppFolio Realm-X”), no Google or Meta advertising connector appeared. We could not enumerate pages 2–17, so this is not a complete negative — verify directly before relying on it.

Myth 6: Google’s Gemini “Developer Assistant” lets you chat with your live account and have it make changes

Per Google’s own docs it is “not a chatbot” but a tool for developers writing code — and “mutating operations (create, update, delete) are generated but never executed.” It drafts change requests; a human executes them.

If you still want to wire it up: the realistic path

Step 1: Decide read-only vs. write access before anything else

For Google Ads, reporting and analysis by chat is achievable today with official tooling. Actual campaign changes — pausing, rebudgeting, bidding — are not, without building or trusting an unofficial write path. This one decision determines every step after it.

Step 2: For Google Ads read-only work, use Google’s official MCP server

Install the official server, authenticate with a developer token, a Google Cloud project ID, and OAuth credentials (client ID/secret or application default credentials), and expect exactly three tools. There is no write tool to configure around.

Step 3: For Google Ads write access, you are building an API client — with a review queue

There is no official MCP write path. You need a Google Ads API client with a Basic token (15,000 operations/day, 5-business-day review) or Standard token (unlimited operations/day, 10-business-day review) per Google’s access levels and permissible use policy, a declared permissible-use category (ad creation and management, reporting, or keyword research), and — if your client provides Creation or Management functionality — Required Minimum Functionality compliance, including letting end users review and edit bid adjustments before they are applied.

Step 4: For Meta, the only sanctioned route we found is the Marketing API itself

With no official Meta MCP server surfacing in our research, the sanctioned route is the Marketing API directly. Start on the default Limited access tier (development only, heavily rate-limited), then apply for Full access via App Review after logging 500+ calls in 15 days with under 15% errors, per the authorization docs.

Step 5: Treat any community MCP server as handing credentials to an unaudited third party

Hundreds exist for both platforms. Anthropic’s own support documentation warns users to “Only connect Claude to servers built and hosted by organizations and applications you trust” and to be aware that the actions Claude takes have no destructive or unintended effects. For Meta tokens specifically, re-read Section 6.a.iv of the Platform Terms first. Also note that Google’s API Terms say you must not misrepresent or mask your identity when using the Google Ads API, and reserve Google’s right to monitor and audit API activity at any time.

Step 6: Check the assistant’s connector ecosystem yourself — do not assume

As of our check, page 1 of Claude’s directory showed no ads connector, and ChatGPT’s first-party connector list could not be confirmed from a primary source (the support article returned HTTP 403 on every fetch). Verify directly.

Step 7: Plan for mandatory human checkpoints on anything touching live spend

Google’s Required Minimum Functionality policy requires human review of bid adjustments for clients providing creation or management functionality, and Google’s own AI dev tool enforces “generated but never executed” for mutations. Treat that as the de facto safety bar for any agent workflow you build or adopt — it is also the design bar we hold AdFlint’s own automation to.

Step 8: Do not mistake a “runs locally” claim for a security sandbox

Google’s Developer Assistant runs generated read-only Python scripts within a “sequestered virtual environment (.venv)” — which isolates dependencies, not credentials, filesystem, or network access. The real safety control in that design is the refusal to execute mutations, not the environment it runs in.

What we could not confirm

Citation-driven research means saying where the trail went cold. Honestly, then:

• We could not enumerate pages 2–17 of Claude’s Connectors directory, so the absence of an ads connector is confirmed only for page 1. The commonly cited “400+ connectors” figure is an inference from 24 per page across 17 pages, not a number stated on the page.
• OpenAI’s support article on ChatGPT connectors returned HTTP 403 on every fetch attempt, so we have no primary-sourced list of ChatGPT’s first-party connectors — and no confirmation of whether any advertising platform appears there.
• Whether Google will add write capability to its official MCP server is unconfirmed. The docs say “Read-only (current release)” and contain no roadmap language either way — phrasing that implies it may change. Re-check before treating read-only as permanent.
• “Meta has no official MCP server” is a negative that cannot be proven conclusively. Our evidence: no MCP mention anywhere in Meta’s Marketing API docs, and no Marketing API server in Meta’s GitHub organisations. Read it as “none found as of July 2026.”
• We could not fetch working primary pages for Google Ads automated rules or Meta Advantage+ campaigns, so this article makes no comparison between MCP-based access and either platform’s own in-product automation.
• The GitHub repo counts (312 and 36) drift daily; they were re-verified live during fact-checking but will differ by the time you read this.
• We did not investigate browser-automation agents that visually operate the Google Ads or Meta web UI rather than calling an API. That is a plausible real-world mechanism outside this article’s MCP-focused scope, and it could change the answer — we simply have no primary-sourced findings on it.

Where AdFlint fits

The pattern across every primary source above is consistent: the official tooling reads, drafts, and recommends, and execution against live spend gets a human checkpoint. That is also the shape of AdFlint’s AI ad management: instead of you wiring a chat assistant to raw API credentials, AdFlint connects to Google Ads and Meta through your own OAuth connection, so you own the ad account outright — no pooled provider account, a distinction that matters more than most people realize (see who owns your ad account). Hard budget caps bound what the automation can ever spend, and the autopilot keeps optimizing after launch rather than stopping at setup. You can read exactly how the pieces fit on how it works and what we do to protect your account on security.

On cost: AdFlint’s self-serve plans are a flat monthly fee with 0% markup on ad spend — your card pays Google and Meta directly — starting at $10/month for up to $100 in monthly ad spend, then $40 up to $1,000, $89 up to $2,500, $109 up to $5,000, $119 up to $10,000, $219 up to $20,000, and $319 up to $30,000, with a 7-day free trial (full ladder on pricing). A separate done-for-you Managed tier starts at a $249/month floor plus a share of managed spend, and is not self-serve. To be clear about our own limits, the same way we ask of the sources above: everything in this article comes from the cited primary documentation, not from us running these MCP servers against live accounts. If you are weighing a chat-assistant setup against a dedicated tool, our comparison of native platform AI vs. third-party tools is the right next read — and if the answer is a dedicated tool, the AI ad manager trial costs nothing for seven days.