Scalewell / Blog

LinkedIn Ads MCP: how to connect Claude or ChatGPT to LinkedIn Ads

LinkedIn's own MCP server is in beta and not yet publicly documented, so most LinkedIn Ads MCP routes are third-party wrappers on the Marketing API; here are the routes in November 2026, what each reads and writes, and how to pick one.

A marketer chatting with an AI assistant that connects through a padlocked gate to an ad campaign dashboard.

A LinkedIn Ads MCP server lets Claude, ChatGPT or another AI client read, and sometimes change, a LinkedIn ad account through the LinkedIn Marketing API. As of November 2026 LinkedIn's own MCP server, which reads and writes ad accounts, is in beta and already in use, but LinkedIn hasn't documented it publicly. Until it opens up, most routes are third-party wrappers, and the real choice is who holds the token.

You know the Monday routine. Campaign Manager export, a pivot table, a slide with cost per lead by campaign, and then the question from the head of demand gen about which of those leads sales actually spoke to. The export can't answer it, and you spend the morning joining spreadsheets by hand. Whichever route you choose, the morning comes back once AI matches those leads to the CRM, and Scalewell builds that into the weekly reporting.

This article is for the paid media lead who wants to ask those questions in a chat instead. It surveys the routes that exist today, confirms from each one's documentation what it reads, writes, costs and needs for authentication, and gives you a decision rule for choosing between them, with no favorite from us. It also covers the LinkedIn-specific traps that the Google and Meta versions of this set-up don't have.

LinkedIn's own MCP server is in beta, beside a read-only ChatGPT plugin and a docs server

LinkedIn has an official MCP server that reads and writes ad accounts. As of November 2026 it is in beta and already in use, but LinkedIn has published no documentation for it, so we can't list its tools, access rules or availability. LinkedIn's two publicly documented AI routes are narrower than their names suggest. The first is a "LinkedIn Ads" plugin inside ChatGPT with LinkedIn listed as the developer. OpenAI's plugin page says it "is currently in beta and supports a limited set of LinkedIn Ads functionality" and "provides read-only access to ad set performance": you can view and search ad accounts and ad sets, ask about performance and break metrics down by date range, ad set or ad, in ChatGPT Work or Codex, with availability depending on your plan and workspace settings. It cannot create, pause or edit anything.

The second is a "Develop with MCP" link, marked "New", on developer.linkedin.com. We opened the destination on 8 November 2026. It is a public page headed "Build LinkedIn API integrations with AI" that tells developers how to point VS Code, Visual Studio, Cursor, Claude Code or Copilot CLI at the Microsoft Learn MCP server (https://learn.microsoft.com/api/mcp) so a coding agent can read LinkedIn's API documentation and write the REST calls for you. It is a documentation server for people building integrations, and it never touches an ad account.

So every LinkedIn Ads MCP, LinkedIn Marketing API MCP or Claude LinkedIn connector you can add to Claude today is a wrapper that someone other than LinkedIn built on the Advertising API. That means you judge each server on how it handles the API, because the API is where the difficulty lives.

A chain of four linked stations from an AI chat window to a server, a keyed gatehouse and the LinkedIn data beyond, with the gatehouse highlighted.
Every route is the same chain: the AI client (Claude, ChatGPT or Cursor) talks to an MCP server (self-hosted or vendor-hosted), which calls the LinkedIn Marketing API through a LinkedIn developer app. The developer app is the highlighted link: whoever owns it owns the approval, the token and the liability.

Every route sits on the LinkedIn Marketing API, so learn its rules first

The LinkedIn Marketing API decides what any MCP server can do, no matter how good the server's tool descriptions sound. Five rules from LinkedIn's own documentation explain most of what you'll see go right or wrong.

RuleWhat LinkedIn's documentation says (checked 8 November 2026)What it means for your AI set-up
Two access tiersDevelopment tier: read any ad account you administer, edit up to 5, create one test account. Standard tier: unlimited edits and account creation, requested through a support ticket with a demo video. SourceAn in-house team with a handful of accounts can live in Development tier. Agencies serving many clients, and every vendor, need Standard.
Member permissions onlyAll scopes are 3-legged, so a named LinkedIn member must consent. r_ads reads accounts, r_ads_reporting reads performance, rw_ads writes. SourceWhoever signs in decides what the AI can touch. The member's ad account role caps the scope (see the read-only trick below).
Token lifetimesAccess tokens live 60 days by default; programmatic refresh tokens, where granted, live 365 days, then the member must reauthorize. SourceA self-hosted server will stop working one day with a 401. Put the date in your calendar.
Monthly versionsEvery call needs a Linkedin-Version header in YYYYMM form; versions are supported for at least one year, and the November 2025 version (202511) sunsets on 16 November 2026. SourceAn abandoned open-source server will break on a date you didn't choose. Check when it was last updated.
Unpublished rate limitsDaily per-app and per-member limits reset at midnight UTC, return 429 when exceeded and are only visible in your app's Analytics tab after a test call. SourceNobody can tell you in advance how many questions a day your team can ask. Make one call, then read the number.

One more rule matters for reporting. The adAnalytics endpoint returns only impressions and clicks unless the request names its metrics, according to LinkedIn's reporting documentation. If an AI answer comes back with suspiciously few metrics, the server probably sent a default request.

Five routes exist, and they differ on who holds the token

Five kinds of route connect an AI client to LinkedIn Ads today, and the practical difference between them is whether you, a vendor or LinkedIn itself owns the developer app and the OAuth token. The table summarizes what we verified on each route's own pages on 8 November 2026.

RouteExamples checkedReadsWritesWho owns the LinkedIn app and tokenCost as of November 2026
LinkedIn's own ChatGPT plugin"LinkedIn Ads" plugin for ChatGPT Work and Codex, developer LinkedIn, in betaAd accounts, ad sets, ad performance by date range, ad set or adNone (read-only)LinkedInIncluded with ChatGPT Work; availability depends on plan and workspace settings
Self-hosted open-source serverdanielpopamd/linkedin-ads-mcp (TypeScript, 25 tools), Nuraveda linkedin-ads-mcp (Python, 11 tools), ZLeventer linkedin-campaign-manager-mcp (TypeScript, 19 read-only tools), CData read-only server (Java, wraps the CData JDBC driver)Accounts, campaigns, creatives, analytics, demographics, lead formsVaries: full create, update and delete; draft-only; or noneYou, through your own developer app and Advertising API approvalFree software (MIT in the first three). CData needs its JDBC driver license or trial.
Hosted connector serviceWindsor.ai, Adspirer, Markifact, InsightfulPipe, KiinPerformance and structure, often blended with other channelsUsually budget, status and targeting changes with confirmation; new campaigns mostly stay in Campaign Manager or land as draftsThe vendor. You sign in with LinkedIn through their OAuth flowFree tiers at Windsor.ai (1 source, 1 account, 30 days), Adspirer (15 tool calls a month) and Kiin (1 account); paid plans from $19 a month (Windsor.ai, annual billing) to $199 a month (Adspirer, 3,000 tool calls), or $65 per ad account per month (Kiin)
General automation MCPZapier MCPCreate reportAudience add and remove, create audience; no campaign create or updateThe vendorIncluded in every Zapier plan; each tool call uses two tasks from your quota
Build your own wrapperClaude Code writing a thin server over the Marketing API, as Anthony Blatner of Speedwork Social documented in April 2026Whatever you buildWhatever you buildYou, with the same approval, token refresh and version header to manage as any open-source serverYour time, plus the model that writes it

Self-hosted open-source servers give you the token and the approval work

Self-hosted servers are free, run on your own machine and need a LinkedIn developer app that you apply for yourself. The three below differ mainly on how far they let the AI write.

danielpopamd/linkedin-ads-mcp advertises 25 tools across campaigns, creatives, audiences, conversions and analytics, including create, update and delete on campaign groups and campaigns. It needs Node 18 or later, takes your client ID and secret as environment variables and runs an OAuth flow with npm run auth. Its README says Advertising API approval takes 1 to 5 business days, which is the author's experience rather than LinkedIn's commitment.

Nuraveda's linkedin-ads-mcp on PyPI, version 0.4.1 released 1 October 2026, is a Python server with seven read tools and four write tools. Every write tool defaults to DRAFT status so nothing goes live by accident. Versions 0.3.x carried a Functional Source License; 0.4.0 onward is MIT.

ZLeventer's linkedin-campaign-manager-mcp is read-only by design across 19 tools, and the author's reason is worth quoting: "Campaign creation has too many failure modes to automate safely; use the Campaign Manager UI." The listing says approval typically takes 2 to 6 weeks and that every call returns 403 until then. Put that next to the 1 to 5 days above and you have the honest picture: LinkedIn publishes no review time, and practitioners report anything from days to weeks.

Hosted connector services hold the token for you and gate the writes

Hosted services remove the developer app, the approval wait and the token refresh, in exchange for a subscription and for trusting a third party with access to your ad accounts.

Windsor.ai covers 240 LinkedIn Ads fields and ships 12 write actions, including enable and pause for campaign groups, campaigns and creatives, plus budget, schedule and targeting changes, each shown to you before it runs; new campaigns still get created in Campaign Manager. The connector is listed in Claude's directory and included on every Windsor.ai plan, including the free one covering one data source and one account for the last 30 days.

Adspirer reads and writes, creates new campaigns paused, cannot delete campaigns and asks for confirmation on budget changes. Markifact is EU-hosted and separates instant reads from approval-gated writes, in its own words "No write executes silently"; its LinkedIn page lists no prices, and its pricing page starts at $79 a month for Pro. InsightfulPipe exposes 28 actions, 18 read and 10 write, lets you connect an account as read-only and starts at $29.99 a month. Kiin runs 95 tools, previews every write as a before-and-after diff and creates new campaigns as drafts.

Pick the route by answering three questions

The right route for your team follows from three questions: who should hold the LinkedIn token, whether the AI should be able to change anything, and how many ad accounts it must edit.

  1. Who holds the token? If your security team won't let a vendor hold OAuth access to your ad accounts, you're self-hosting or building, and applying for the Advertising API yourself. If they're fine with it, a hosted service saves the approval wait and the refresh chore; read its data handling terms first, because you are granting rw_ads or r_ads on your accounts to their app.
  2. Read, or read and write? For reporting and analysis in a chat, pick a read-only route or connect as a read-only user and stop worrying. If you want the AI to pause campaigns or move budget, pick a route whose write gating you've watched work, and read our view on where humans stay in the loop on ad spend before you let it near a budget field.
  3. How many accounts need editing? Development tier edits up to five ad accounts you administer and reads without limit, so one brand with five accounts or fewer fits inside it. An agency editing 20 client accounts needs Standard tier, which means the support ticket, the demo video and a review with no published timeline, or a hosted vendor who has already been through it.
A forking path with three splits leading to four different destinations, standing in for the choice between self-hosted and hosted routes.
Three questions, four routes. Who holds the token: you or a vendor? Read only, or read and write? How many accounts need editing: five or fewer, or more? Self-hosted read-only suits a team that keeps the token and only reports; self-hosted writes fit inside Development tier for up to five accounts; a hosted service suits a team happy to let a vendor hold the token; more than five accounts to edit means Standard tier or a hosted vendor.

The self-hosted route takes six steps and one wait of unknown length

The self-hosted route is a developer app, an Advertising API application, an OAuth flow and a client configuration, with a wait in the middle that LinkedIn puts no number on. The steps follow LinkedIn's quick start and key concepts pages and the README of the server you pick.

  1. Create a LinkedIn Page and a developer app. LinkedIn requires an associated Page before you create an app in the Developer Portal. Use a company identity, because LinkedIn's restricted use cases page says "Do not create fake profiles to manage ad accounts or use individual profiles to manage multiple customers' accounts."
  2. Apply for the Advertising API on the Products tab. Complete the access form, describe an internal reporting or campaign management use case for accounts you administer, and wait for the approval email.
  3. Add your ad accounts to the app. Copy the nine-digit account ID from Campaign Manager, open your app, go to Products, click View Ad Accounts, choose Add Ad Account and save. Only accounts added here can be managed through the API in Development tier, and you can add up to five.
  4. Set the redirect URL and run the OAuth flow. On the Auth tab add the redirect your server expects (danielpopamd's README and Speedwork's tutorial both use http://localhost:3000/callback), put your client ID and secret in the server's environment variables, and run its auth command. Sign in as the member whose ad account role you want the AI to inherit, and grant only the scopes you need: r_ads and r_ads_reporting for reading, rw_ads only for writes.
  5. Register the server with your client. In Claude Code the documented form for a local server is claude mcp add --transport stdio <name> -- <command> [args], with -e KEY=value for environment variables and /mcp to check status, per Claude Code's MCP documentation. In Claude Desktop, open the Claude menu, then Settings, Developer and Edit Config, add the server under mcpServers in claude_desktop_config.json, and fully quit and restart the app, per the MCP user quick start.
  6. Ask one question and reconcile it. Ask for last week's spend by campaign and compare it to Campaign Manager before you trust anything else; spend and clicks should match to the cent and the click. Then open your app's Analytics tab in the Developer Portal to read the rate limit that the test call revealed.

The hosted route is two sign-ins, and the decision sits in Claude's connector settings

A hosted route is a vendor sign-in, a LinkedIn sign-in and one connector added in Claude, and the thing to get right is who in your organization may add it.

  1. Create the vendor account and connect LinkedIn. Authorize LinkedIn through the vendor's OAuth screen and pick the ad accounts it may see. If the vendor offers a read-only connection, as InsightfulPipe does, take it until you have a reason not to.
  2. Add the connector in Claude. Anthropic's custom connector guide puts the path at Customize, then Connectors, then "+ Add", then "Add custom connector", where you paste the vendor's server URL, or you click Connect on the vendor's listing in Claude's directory. Custom connectors are available on Free (limited to one) and on every paid plan; on Team and Enterprise plans only an owner can add them, under Organization settings, then Connectors.
  3. Read the warning and mean it. Anthropic's guide says "Custom connectors allow you to connect Claude to arbitrary services that have not been verified by Anthropic." OpenAI's MCP documentation says the same of ChatGPT and adds that ChatGPT requires manual confirmation before write actions. Then run the same spend-by-campaign reconciliation before anyone presents a number from the chat.

In ChatGPT the server URL becomes a plugin, and full support needs Business or Enterprise

ChatGPT adds a custom MCP server as a plugin, and as of November 2026 OpenAI's help center limits full support to "ChatGPT Business and Enterprise/Edu customers on ChatGPT web". Pro users "can connect custom MCP servers with read/fetch permissions", and no Developer Mode toggle is needed. The steps follow OpenAI's Add custom MCP server guide and plugin quickstart.

  1. Check who may add it. On Business only admins and owners can create and publish these plugins. On Enterprise or Edu an admin can also grant a member the "Create plugins with MCPs" permission.
  2. Create the plugin. Open chatgpt.com/plugins, select the plus button, then "Add custom MCP server". Enter a name, paste the vendor's server URL under Connection, set the authentication the vendor documents, tick "I understand and want to continue" on the risk warning, then "Create as a plugin".
  3. Install it. Install it from your personal plugins and complete the vendor's sign-in if its server uses OAuth.
  4. Call it in a Work chat. Switch the tab from Chat to Work, type @, pick the plugin and ask for last week's spend by campaign. By default ChatGPT asks you to confirm any tool the server has not marked with the readOnlyHint annotation.

ChatGPT connects only to remote servers over SSE or streaming HTTP, so a self-hosted stdio server needs an HTTP transport and a public HTTPS endpoint or OpenAI's Secure MCP Tunnel.

If you are doing this for Google Ads or Meta at the same time, the platform-side rules differ a lot, and we cover them in Google Ads MCP and Meta Ads MCP.

Signing in as a VIEWER makes any server read-only, whatever it claims

The LinkedIn ad account role of the member who signs in caps what every tool can do, and LinkedIn's campaign documentation states that the VIEWER role is "read-only, even with rw_ads scope". That is the most useful sentence in this article for anyone nervous about letting an AI near an ad account.

It works like giving a contractor a key that only opens the meeting rooms. Run the OAuth flow as a member with the VIEWER role on the accounts in question, and it no longer matters whether the server exposes delete_campaign or how well its confirmation prompt works, because the API refuses the write. The role list runs VIEWER, CREATIVE_MANAGER, CAMPAIGN_MANAGER, ACCOUNT_MANAGER and ACCOUNT_BILLING_ADMIN, so you can grant exactly the write level you want.

When you do want writes, prefer servers that default to draft or paused. LinkedIn's campaign statuses include DRAFT, described as "still being edited and not eligible for serving", and a draft is a proposal you can read in Campaign Manager before anything spends.

Most failures come from the API's rules, stale servers and misread numbers

Most problems with a LinkedIn Ads MCP come from the API's own rules, from servers nobody maintains and from reading an AI's summary as if it were a Campaign Manager export.

The token expires and nobody notices. Access tokens last 60 days. A self-hosted server that answered questions in September goes quiet in November with a 401, usually on the morning of a reporting deadline. Hosted vendors handle this; if you self-host, someone owns the renewal, and the same person owns checking the version header before LinkedIn sunsets it.

The reporting request asked for the defaults. If a server sends no fields list, LinkedIn returns impressions and clicks and nothing else, and the AI will cheerfully report a click-through rate while telling you it couldn't find spend. The fix is a server that names its metrics, and a reconciliation habit on your side.

Demographic numbers don't match. LinkedIn says member demographic metrics such as job function and company size can lag 12 to 24 hours, are approximate to protect member privacy, return only the top 100 values per creative per day and drop values with fewer than three events. A chat answer about which job titles clicked is a sketch.

The server scrapes instead of calling the API. Some community LinkedIn servers for profiles and posts drive a browser against the logged-in site rather than calling an API, as Tony Lewis at MCP Bundles documented in April 2026. None of the Ads servers above do, but LinkedIn's User Agreement section 8.2 prohibits "bots or other unauthorized automated methods" and any "scripts, robots or any other means or processes" to scrape the services, so check before you install anything that asks for your LinkedIn password instead of showing an OAuth screen.

You expected pipeline and got clicks. The Marketing API does carry a revenue attribution finder that reports opportunities and revenue won, but LinkedIn's reporting documentation says it only works for accounts that have connected Salesforce, Dynamics 365 or HubSpot through Business Manager, and of the servers we checked only InsightfulPipe advertises it, with a note that it still needs the CRM connected first. For the question that started this article, which leads sales spoke to, the ad data needs your CRM beside it, which is a separate connection and the subject of AI lead scoring for demand gen.

Who should wait: a team with nobody willing to own the token renewal and the version check, a team whose only goal is to let an AI create campaigns unattended, and any agency hoping Development tier will cover 20 client accounts. In each case the set-up works for a month and then quietly fails you.

Where to start

Start with a read-only connection under a VIEWER role, ask for last week's spend by campaign, and reconcile it to Campaign Manager before you ask anything harder. If your team wants help choosing a route and setting the method up around how you already report, that is the kind of work we do at Scalewell.

Frequently asked questions

Do I need LinkedIn Marketing API access to use a LinkedIn Ads MCP?

It depends on the route. Self-hosted and home-built servers need your own developer app with the Advertising API approved, which starts in Development tier. Hosted services such as Windsor.ai, Adspirer, Markifact, InsightfulPipe and Kiin hold their own approval, so you only sign in with LinkedIn and grant their app access to your ad accounts.

Can Claude create LinkedIn campaigns through an MCP server?

Some servers expose campaign creation, and the LinkedIn Marketing API allows it in Development tier for up to five accounts you administer. In practice several servers refuse or default to draft: ZLeventer's is read-only on purpose, Nuraveda's creates drafts, and Windsor.ai leaves new campaigns to Campaign Manager. We'd keep campaign creation in the UI for now.

How long does LinkedIn Advertising API approval take?

LinkedIn publishes no review time and says it reserves the right to select partners at its discretion. Server authors report very different experiences: one README says 1 to 5 business days, another listing says 2 to 6 weeks, and one claims approval is automatic for accounts with an active Campaign Manager. Plan for weeks and be pleased if it's days.

Is it safe to give an AI client write access to my LinkedIn ad account?

It's safe when the role and the server both limit the blast radius. Sign in as a member whose ad account role matches what you want the AI to do, since VIEWER is read-only even with the rw_ads scope, and prefer servers that preview each write and default new campaigns to draft or paused. Both Anthropic and OpenAI warn that custom connectors are unverified by them.