The short answer
The best MCP servers for Meta ads split into two families that answer different questions. Account servers (Meta's official Ads AI Connectors, Pipeboard, open-source servers) read and change your own campaigns. Research servers (Ad Radar, Foreplay, GetHookd, Atria) search other brands' ads. Meta's own server is free in open beta and the default for your account; for research, pick by scope and price. Most teams end up connecting one of each.
In this article5 sections
The best MCP servers for Meta ads come in two families
Picking the best MCP servers for Meta ads starts with one question: whose ads do you want Claude to see? An MCP server is a connector that gives an assistant like Claude a set of tools on a real system. For Meta ads, those systems fall into two groups that don't overlap.
- Account servers connect to your ad account. They pull reports, create campaigns and edit ad sets. They can't see what other brands run.
- Research servers connect to a database of other brands' ads. They search competitors, open transcripts and compare patterns. They never touch your account.
A media buyer asking "why did CPA rise on Tuesday?" needs the first. A creative strategist asking "what are the scaling hooks in skincare this month?" needs the second. Here is the short list, verified on each vendor's own pages on October 12, 2026.
| Server | Family | Connects to Claude via | Price |
|---|---|---|---|
| Meta Ads AI Connectors | Your account | Hosted URL, business login | Free in open beta |
| Pipeboard | Your account | Hosted URL or self-hosted | Free tier, then $29.90-$199/month |
| GoMarble Facebook Ads MCP | Your account | Self-hosted, your access token | Free, open source |
| Ad Radar | Competitor research | Hosted URL, OAuth | $49 every 4 weeks (launch price) |
| Foreplay | Competitor research | Hosted URL, OAuth | Workflow plan, $175/month |
| GetHookd | Competitor research | Hosted URL, API token | Annual plans from $19 or $32/month |
| Atria | Research plus your accounts | Hosted URL, OAuth | Core $129/month, billed annually |
| Community Ad Library servers | Competitor research | Self-hosted | Free code, paid data key |
If the protocol itself is new to you, MCP for marketers covers it in five minutes.
Account servers: Claude inside your ad account
Meta Ads AI Connectors (official)
Meta opened its own connectors in open beta on April 29, 2026: a hosted MCP server and a CLI that share the same API (PPC Land). The server URL is https://mcp.facebook.com/ads. In Claude you add it as a custom connector, sign in with Facebook Login for Business and choose which businesses and ad accounts Claude may reach. No developer app, no token to copy.
It reads reporting and benchmarks, and it writes: campaigns, ad sets, ads, catalogs and dataset diagnostics. Objects it creates land paused by default, so a person has to switch them on (MCP Directory). It is free during the beta, and Meta is rolling it out gradually, so some accounts see it later than others.
Best for: anyone who runs Meta ads and wants Claude to read or build in their own account. It is the default choice for the account family.
Pipeboard
Pipeboard ran a Meta Ads MCP before Meta had one, and its site now lists Google, TikTok and Reddit Ads in the same connector. The Meta server is on GitHub as meta-ads-mcp, under the Business Source License, set to become Apache 2.0 in 2029. You can self-host it with your own Meta developer app, but the project recommends its hosted version, which you add to Claude as a connector and log into with a Pipeboard account.
Pricing: a free plan with 30 AI executions a week and 2 ad accounts, then Pro at $29.90, Premium at $99 and Agency at $199 a month, with more accounts, team seats and bulk operations as you go up.
Best for: agencies that want team access controls, bulk creative uploads and several ad platforms behind one connector. If you only need one Meta account read and edited, Meta's free server now covers that.
GoMarble Facebook Ads MCP (open source)
GoMarble's server is MIT-licensed Python you run on your own machine. You pass it a Meta user access token with ads_read permission and point Claude Desktop at it. Its 20-plus tools are mostly reads: accounts, campaigns, ad sets, ads, creatives, insights and the change history.
Best for: developers who want to see every line of code that touches their account, or who need a read-only setup they control.
Research servers: Claude inside other brands' ads
Your account tells you what worked for you. It can't tell you what is working for the other brands in your niche. That's the job of this family.
Ad Radar
Ad Radar is a research board of Meta ads with $50k or more in estimated spend, each tagged on ten dimensions: hook pattern, hook mechanism, narrative angle, awareness, proof, production format, video type, landing page type, niche and CTA. On our October 9, 2026 snapshot that is 2,443 ads from 749 brands, with a median estimated spend of $223K.
2,443
winning Meta ads on the board
749
brands running them
$223K
median estimated spend
6
tools Claude can call
The MCP server gives Claude six tools:
| Tool | What it does for you |
|---|---|
| search_ads | Searches ads by niche, hook, angle, format or a line from the video |
| get_ad | Opens one ad's full record: copy, script, signals, days live, variants |
| winning_patterns | Shows which hooks and angles win more often than the rest of the board |
| advertiser_hooks | Shows the hook mix one brand is running |
| get_swipe_prompt | Returns a ready brief to remake a winner for your product |
| list_taxonomy | Lists the valid tag values, so Claude filters with the right words |
Setup in Claude Code is one line; on claude.ai or Claude Desktop you add the same URL as a custom connector and sign in with OAuth:
claude mcp add --transport http ad-radar https://app.ad-radar.dev/api/mcpThe server is included in the Membership plan, $49 every 4 weeks at the launch price (list $99). The limits are on purpose: Meta only, no TikTok, and it does not connect to your ad account. A full session is in Claude for marketing research.
Best for: strategists and buyers who want Claude to work from ads that already carry real budget, ranked by estimated spend, rather than from every ad in the library.
Foreplay
Foreplay's MCP connects Claude, ChatGPT and other clients to its discovery library of 200M+ ads across Facebook, Instagram, TikTok, YouTube and LinkedIn, plus your own boards, swipe file and tracked brands. You add https://public.api.foreplay.co/mcp as a custom connector and sign in. Access starts on the Workflow plan, $175 a month or $149 billed annually; the $59 Basic plan doesn't include it. Calls use the API credits in your plan, with no separate per-query charge.
Best for: teams that research across several platforms and already keep their swipe file in Foreplay. Our side-by-side is in Ad Radar vs Foreplay.
GetHookd
GetHookd's server exposes a long tool list (search, brand tracking, boards, swipe files, ad cloning) over a library it describes as 22M+ live Meta ads. It authenticates with an API token you create in its settings (setup guide). Some tools are free to call and others spend credits from your plan. API and MCP access come with any annual plan, from $19 or $32 a month billed yearly depending on which of its two price lists you see, or with the two higher monthly plans at $79 and $129.
Best for: buyers who want the biggest Meta library at the lowest entry price and are fine managing a token and credits. See Ad Radar vs GetHookd.
Atria
Atria sits between the families. Its MCP server searches Atria's ad and brand libraries and also reads your boards and connected ad accounts, transcribes ad videos and generates images. You add https://api.tryatria.com/mcp as a connector and sign in with OAuth. On its pricing page, Core is $129 a month billed annually with 600 MCP calls a month, and Plus at $479 has no call limit.
Best for: brands that want research and their own account data in one connector and have the budget for it. Compare in Ad Radar vs Atria.
Community Ad Library servers
Open-source servers such as facebook-ads-library-mcp (MIT) look up a brand's ads in the public Meta Ad Library and can analyze images and video. The code is free; the data comes through a third-party API key you pay for. Like the Ad Library itself, they show what is running, not what is spending: Meta publishes spend only for political and social issue ads.
Best for: a quick look at one brand's live ads from inside Claude, when you don't need ranking or tags.
How to choose: match the server to the question
| Your question | Server family | Start with |
|---|---|---|
| Why did my CPA move this week? | Account | Meta's official server |
| Build next week's test campaign for me | Account | Meta's official server or Pipeboard |
| Which hooks are scaling in my niche? | Research | Ad Radar |
| What is this competitor running on TikTok and Meta? | Research | Foreplay or Atria |
| Show me everything one brand has live | Research | GetHookd or a community server |
Three checks before you connect anything:
- Start read-only. A server that can create campaigns can create the wrong one. Keep budget and status changes behind your approval, one account at a time.
- Ask which tools it called. If an answer looks off, have Claude list the tools and filters it used. Most bad answers are a bad filter.
- Treat outside text as data. Competitor ad copy can contain lines that read like instructions. Claude should quote it, not follow it.
Connect one of each
The two families work best together, because the useful questions cross the line. "Our UGC ads fatigue after two weeks; what are competitors doing differently?" needs your account data and the market.
A session that uses both, in plain prompts:
1. From my Meta account: list my top 10 ads by spend in the last 30 days
with CPA and days live. Tag each one by hook type and format.
2. From Ad Radar: search_ads in my niche, active, sorted by score,
limit 25. Then winning_patterns for video.
3. Compare: which hook types and formats are carrying competitor
budgets that I am not testing? Name the ads behind each gap.
4. Pick the strongest gap and call get_swipe_prompt on the best donor.
Label every competitor spend figure as estimated.Step 1 runs on the account server, steps 2 to 4 on the research server, and Claude does the comparison. If you repeat a session like this every week, save it as a prompt or as one of the Claude skills for Meta ads; Claude prompts for Facebook ads has more in this shape, and Claude for media buying covers what to hand Claude and what to keep.
Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other. How the board is built.