Meta Ads MCP and LinkedIn Ads in Claude and ChatGPT: the ways to connect your ad accounts

We read every page linked below on 23 October 2026. Meta changes its connector often (its tools page shows an update on 16 October 2026), so treat a fact older than a month as a reason to check again. Where a page does not say something, this guide says "the page does not say".

What is an ads MCP server, and what can an assistant do with it?

An ads MCP server is a program that lets an AI assistant such as Claude or ChatGPT call functions on an ad platform. MCP stands for Model Context Protocol, which its own site calls "an open-source standard for connecting AI applications to external systems". The server lists its tools. The assistant picks a tool when your question needs one and shows you the result.

Each tool does one job. Examples are listing ad accounts, returning results by campaign, searching an ad library or creating an ad. The server decides which jobs exist, so one server may only read while another can change campaigns. The MCP specification calls tools "model-controlled" and says "there SHOULD always be a human in the loop with the ability to deny tool invocations". How closely a product follows that line differs.

There are six ways to let an assistant see your ads:

• Meta's own ads MCP server, for Meta ads only.
• A community server, which a developer runs or hosts.
• A hosted connector from a data or reporting vendor.
• A file: an export or screenshots that you upload to the chat.
• A creative analytics tool with its own MCP server, Nimblet among them.
• An ad library server, which reads other advertisers' public ads and not your account.

An ad library server answers "what do my competitors run". An account server answers "what did my ads do". Pick by that question.

What does Meta's own Ads MCP server offer?

Meta hosts an official ads MCP server at https://mcp.facebook.com/ads. It reads results, and it also creates, edits and deletes ads, ad sets, campaigns and custom audiences. Meta announced its ads AI connectors as an open beta on 29 April 2026. On 16 July 2026 it opened the server to developers who use their own Meta app (Meta announcement, Meta developer post). We found no page that says the beta has ended.

Meta's developer overview lists seven tool groups: reporting, ad creation and management, catalogs, signals and datasets, help and troubleshooting, A/B tests and conversion lift studies, and activity logs. Meta's Business Help article adds custom audiences (create, update and delete). It says the agent can create single image, single video, carousel and product media carousel creatives, and upload image or video files.

The tools page names the tools. These are some of them:

• ads_get_ad_entities returns campaigns, ad sets and ads with results.
• ads_get_creatives, ads_get_ad_images, ads_get_ad_videos and ads_get_ad_preview cover creatives.
• ads_library_search searches the public Meta Ad Library.
• ads_create_ad and ads_update_entity write.
• ads_activate_entity turns a paused item on, so that it starts to spend budget.

Meta gives no tool count. Its developer post says it is rolling out access to tools gradually and tells you to ask your agent which tools you have. This guide prints no count. The test prompt below asks your assistant for that list.

There are two ways in. A business user without a Meta app follows the Business Help article and connects from Claude, ChatGPT, Claude Code or Perplexity. A developer adds the use case "Create & manage ads with ads MCP server" to their own Meta app. Sign-in is OAuth through Facebook Login for Business, a user token or a system user token (setup page). The minimum permissions are ads_mcp_management plus ads_read or ads_management. Only Employee-role system user tokens work. An agency that manages other businesses' data must submit its app for review and request Advanced Access on ads_mcp_management (developer post).

No Meta page we read states a price. Two third-party articles call the server free. One is from Passionfruit. The other, from a creative analytics vendor, adds "during the public beta". Meta does not say so.

Is there a LinkedIn Ads MCP server?

We found no MCP server run by LinkedIn on 23 October 2026. LinkedIn's Marketing API hub on Microsoft Learn lists no MCP page, and a search of LinkedIn and Microsoft Learn pages found none either. A vendor of a hosted LinkedIn connector wrote on 10 June 2026 that "there is no official one". LinkedIn may add one, so check the hub before you build.

A live connection to your LinkedIn ad data goes through LinkedIn's Advertising API, and that API has access rules. LinkedIn's access tiers page (dated 2026-04-01) says access "is reserved for applications building supported use cases". The Development tier allows unlimited reads on ad accounts you administer and edits on up to 5 ad accounts. LinkedIn "reserves the right to review applications and select partners at its discretion". The same page warns that Marketing version 202511 sunsets on 16 November 2026, so anything built on the API needs a version check.

What follows from this:

• A community server that you run yourself needs your own LinkedIn developer app with Advertising API access.
• In our reading, a hosted connector uses the vendor's own API access, and you connect your ad account to the vendor's service.
• A file upload needs no API access.

The table in the next section lists the community servers and hosted connectors we opened, with what each reads, whether it can write and the page we read.

What can each setup see at creative level?

Most pages we read describe numbers and text fields. For Meta's server and for most hosted connectors, they do not say whether the assistant gets the image or the video, so treat that as unknown until you test it. Three setups say more. Pipeboard, a community Meta server, says its tools return creative details with images and URLs. A file upload gives Claude the images you attach. Nimblet passes a stored written read of each ad.

The table lists the ways to connect your own ad accounts. "The page does not say" means unknown.

Some creative analytics and ad library tools also run an MCP server. They read public ads, their own database or both, and they answer "what do competitors run". Some also connect your account. Platforms, access and write rights differ by tool and by plan, so read each vendor's own page. For a side-by-side view, see Nimblet vs Motion, Nimblet vs Foreplay and Nimblet vs Segwise.

Meta's tools page names image, video and preview tools and does not describe what they return. One third-party article says Claude "cannot see your ad images or videos through the MCP". That article describes no test. So the question is open for Meta's server and for every row marked "the page does not say". Run the test below on one ad. It takes a few minutes and works on any row.

You are connected to [connector name] for [Meta or LinkedIn] ad account [account name]. Run a test on one ad. Use only the connector's tools. Do not guess. Do not change anything.

Step 1. List every tool the connector gives you. Use the exact tool names. Mark each tool "read" if it only returns data. Mark it "write" if it can create, edit, pause, activate, upload or delete anything.

Step 2. Pick the ad named [ad name]. Return its headline, its primary text and its call-to-action button exactly as the connector gives them. For any field the connector did not return, write "not returned".

Step 3. Say what the connector gave you for the picture or video: the file itself, a link, an ID only, or nothing. Count only what you can read yourself, not what the chat window may show me. If you have the file, describe in one sentence what a viewer sees first, quote any words printed on the image and name its two main colors. If you have a link, an ID or nothing, write "I cannot see the creative" and do not describe it. Never infer the look of the ad from its text or its name.

Step 4. Finish with a table with one row per field: headline, primary text, call-to-action, image, video, landing page URL, run dates, spend, results. Mark each field "returned" or "not returned".

Do not call any write tool.

Open the same ad in Ads Manager or Campaign Manager and compare the headline and text with the answer. If they differ, stop and do not rely on the setup. Then check the words and colors it named for the picture. If they are wrong, the assistant did not see the image, whatever it said.

Keep a record for each connector you test. A record with a date and a page link is what you paste into the team channel before someone adds write access.

Connector record, checked on [date]

Name: [connector name]
Server URL: [URL]
Run by: [vendor, open-source author, or "us"]
Platforms: [Meta, LinkedIn]
Read tools it gives: [tool names from the test]
Write tools it gives: [tool names from the test, or "none"]
Creative test result: [text only, link only, or image seen]
Sign-in: [OAuth or token]. Permissions it asks for: [list]
Account role used for the connection: [Analyst, Viewer, or other]
Set up by: [name]
Page we read for each fact: [URL]
Next check: [date, one month ahead]

How do you connect Claude or ChatGPT?

In Claude, add the server as a custom connector under Customize, then Connectors. In ChatGPT, add it as a custom MCP server under Plugins, on the web, in a Business, Enterprise or Edu workspace. Claude reaches the server from Anthropic's cloud, so the address must be reachable from there. ChatGPT asks for a public HTTPS address, or a Secure MCP Tunnel for a private server. Meta's Business Help article points to the vendors' own guides and lists Claude, ChatGPT, Claude Code and Perplexity as supported agents.

Claude, from Anthropic's custom connector guide (the steps for Pro and Max):

1. Open Customize, then Connectors.
2. Click Add, then Add custom connector.
3. Enter a name and the server URL, then click Continue.
4. Check the sign-in settings that Claude found, then click Continue.
5. Choose how people sign in, then click Add.
6. In a chat, click the plus button, choose Connectors and switch the connector on for that chat.

Custom connectors work on Free, Pro, Max, Team and Enterprise. Free allows one. On Team, only Owners and Primary Owners add a custom connector, under Organization settings, then Connectors. Members then click Connect. On Enterprise, people with a custom role that includes "Manage access to Libraries" can also add one.

ChatGPT, from OpenAI's help article:

1. Go to Plugins, click the plus button and choose Add custom MCP server.
2. Enter a name. Under Connection, enter the server's public HTTPS address.
3. Set the sign-in method. For Meta's server with your own Meta app, choose OAuth and enter your Meta app ID as the OAuth client ID (Meta setup page).
4. Tick the risk warning, create the plugin and read the tools it found.
5. In a new chat, type @ and pick the plugin.

On a Business workspace, only admins and owners can create and publish a custom MCP server. On Enterprise and Edu, admins can allow selected members. OpenAI says custom MCP servers are web only. It says Pro users can connect them with read and fetch permissions only, and that full MCP is for Business, Enterprise and Edu. Menu names change. OpenAI's article now says you do not need to turn on Developer Mode first, and Meta's setup page still links a developer mode guide for ChatGPT. Follow the menu you see.

For Claude Code, Meta's setup page gives this command: claude mcp add --transport http --client-id <META_APP_ID> meta-ads https://mcp.facebook.com/ads.

For Nimblet, the steps are the same. Use https://app.nimblet.io/api/mcp as the server URL and sign in with your Nimblet login. Details are on the Nimblet MCP page.

For the file route, export a report from Ads Manager or Campaign Manager and attach it to the chat. Screenshots of the ads give the assistant the pictures that an export lacks.

Run the test prompt first. When it passes, the prompts in AI prompts to analyze B2B ads and the workflow in writing B2B ads with Claude or ChatGPT work on top of the connection.

What should you check before you give an assistant write access?

Before you give an assistant write access, decide whether you need it at all, learn what the connector may change, set limits on the platform and in the assistant, name who can connect and plan how you undo a change. If you only need to read, keep read access and stop here. The facts below come from the vendors' own pages. The template that follows is the checklist.

What the connector may change. Meta's tools page lists tools that create and update campaigns, ad sets and ads, turn paused items on, and delete custom audiences. It says "your AI client asks for confirmation before activation", so by Meta's wording the confirmation sits in your AI client. In the text we read, it describes no confirmation step for deleting a custom audience or for removing users from a customer list.

What Meta lets you limit on the platform side. If you have full control of a business portfolio, open Settings in Meta Business Suite, then Integrations, then Ads MCP server. You can block actions such as creating campaigns, block all budget changes, or block budgets above an amount you enter. Meta says the server enforces these rules. A partner with access to your ad account can also set rules, and they apply only to the partner's own employees. If you do not see the menu, you do not have the feature yet (Business Help). Meta's developer post gives developers the same advice: start an account in a read-only or approval-required posture and widen permissions as confidence grows. The Business Help article also says "All ads are paused by default until you set them live" and "Any actions taken on your behalf require your authorization through the AI agent."

What the assistant limits. In Research, Anthropic says "Claude can invoke tools from your connectors automatically without further approval", and tells you to disable write-capable tools before you use it. It also tells you to click "Allow always" only for servers and tools you trust to run unsupervised. OpenAI says ChatGPT "may ask for confirmation" for write actions and that "some especially risky actions may be blocked". Deep research can use custom apps for read and fetch actions only. After an admin approves a server, ChatGPT keeps a frozen copy of its tools, and later changes to the server do not apply until an admin reviews them.

Text from outside your team. Anthropic warns: "Malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions." OpenAI warns about prompt injection from untrusted servers. In our reading, the same risk applies to text that comes from outside your team, such as a competitor ad or a lead form answer. An assistant that can write could follow an instruction hidden in it. Meta's server has both kinds of tool: ads_library_search reads other advertisers' ads and ads_update_entity writes. Keep write tools off when the assistant reads such text.

Roles as a second lock. Meta's ad account roles include Analyst, which is view-only. LinkedIn's Campaign Manager roles include Viewer, which can view performance but cannot create or edit campaigns or ads (Meta roles, LinkedIn roles). In our reading, a connection made with such a role should not be able to change the account, because the platform decides what the role may do. We have not tested this. Test it with a harmless request before you rely on it.

Scopes and calls. A scope shows what a tool may do. The tool's calls show what it does. Ask each vendor which calls it sends. Nimblet is an example. Its sign-in asks LinkedIn and Meta for permissions that can write, and its code makes read calls only to your ad accounts. It never creates, edits or pauses a campaign.

Ban risk. Meta writes: "It's safe to connect your ad account through Meta's ads MCP server, and it's unlikely that an account would be banned simply for using it." Meta ties bans to breaking its advertising standards. The Meta pages we read say nothing about third-party servers. A creative analytics vendor wrote on 20 July 2026 that third-party, non-approved MCP servers have been getting ad accounts banned since 2025. We found no Meta page that says so. Passionfruit reports the same rumor and says Meta has confirmed no link to specific tools. Treat it as an unproven vendor claim.

Write access check for [assistant name] on [platform] account [account name]

Owner: [name]    Date: [date]    Connector: [name and URL]

1. Need: [the one change we want, for example "build paused draft ads"]. If we only need to read, we stop here and keep read access.
2. Read first: we ran the connector on read access for [number] days. We compared its answers with [Ads Manager or Campaign Manager] on [date]. Result: [matched, or list the gaps].
3. Permissions: the connector asks for [list]. Tools each one allows: [list]. We removed [permission] because [reason].
4. Platform limits: in Meta Business Suite, Ads MCP server rules are set as create campaigns [Allowed or Blocked], edit or set budget [Allowed or Blocked], budget cap [amount]. LinkedIn: limit we set on the platform [limit, or "none found"]. Role used for the connection [Analyst on Meta, Viewer on LinkedIn, or other].
5. Assistant limits: write tools are [off or on] in [Claude or ChatGPT]. "Allow always" is [off]. Research or deep research is [off for this connector].
6. Who can connect: [named people]. A shared login is not allowed.
7. Go live: nothing the assistant creates runs until [name] switches it on in [Ads Manager or Campaign Manager].
8. Audit and undo: we read [the Meta activity log, or the change history of the platform we use] every [day of the week]. To undo a change we [steps].
9. Sign-off: [name], [date]. Next review: [date, one month ahead].

Run this once per connector and per ad account. Review it each month, because Meta changes the connector and the tool list.

When does a hosted tool beat a do-it-yourself setup?

A do-it-yourself setup wins when you run Meta only, when you want the assistant to build paused drafts and edit campaigns, and when you ask a few one-off questions. A hosted tool wins when LinkedIn carries budget and you have no developer app, and when you want the same answer next week from a stored read of each ad. The full build-or-buy decision is on the Nimblet versus a do-it-yourself setup page.

Do it yourself when:

• You run Meta only. Meta runs the server itself, with its own rules for what an agent may change, and no Meta page we read states a price.
• You have a person who keeps tokens, permissions and API versions in order.
• Your questions are about numbers, and you check the first answers against the platform.

Use a hosted data connector (Coupler.io, Two Minute Reports, Reportei, Windsor.ai) when you want LinkedIn or Meta numbers in a chat without a developer app. The pages we read describe numbers and creative metrics. They do not say that the assistant sees the ad.

Use a creative analytics tool when the question is about the ad itself. Check which platforms it reads, whether it reads your own account or only public ads, and whether it can write. Some of these tools cover Meta only.

Nimblet reads LinkedIn and Meta, and only those two, in one connection. It stores a written read of each ad with recent spend and uses it with your brand context, your documents and the competitors you add from the LinkedIn and Meta ad libraries. It does not create or edit campaigns. If you need to change campaigns from a chat, use Meta's server for that part. Claude and ChatGPT can both use more than one connector in a chat, so you can connect both.

Whichever you choose, take a number that goes to a client or to finance from Ads Manager or Campaign Manager. Use the assistant to find what to look at.

To try Nimblet in Claude or ChatGPT, start at the Nimblet MCP page. MCP needs a paid plan, and pricing lists the plans.