MCP Ads

MCP Ads: Google Ads MCP server + Meta Ads MCP server

MCP Ads, the ads MCP server by Roy Azran. Hosted at mcp-ads.com, open source here.

Website · Docs · Pricing · MCP Registry listing (com.mcp-ads/mcp-ads)

[License: MIT] [CI]

Open-source (MIT) MCP server for Google Ads, Meta Ads (Facebook & Instagram), GA4, Search Console, Tag Manager, TikTok, LinkedIn, Microsoft Ads and more. Works with Claude, Claude Code, ChatGPT, Cursor, Codex, Gemini CLI, Grok, or any MCP client. Read performance, audit wasted spend, build reports, and create or change campaigns. Everything it creates starts paused until you say otherwise.

3,363 registered actions across 21 tool families in the hosted build (September 2026), exposed through one dispatcher per family so your agent loads about two dozen tool definitions instead of thousands. This repo ships 18 of those families; see Tool categories.

How it compares

Facts as of September 2026, taken from each product's public pages. Plans and coverage change, so check their site before deciding.

Use Google's or Meta's server if you only need one platform and want the vendor's own code. Use this one if you want one server across platforms, write access to Google Ads, or the option to run the whole thing yourself.

Choose your path

☁️ Cloud — connected in 2 minutes

Go to mcp-ads.com, sign in, OAuth your ad accounts in the browser, and paste one URL into Claude. Done.

• No API applications, no developer tokens, no app reviews — the platform approvals (Google Ads developer token, Meta App Review, and the rest) are already in place
• $49/month flat for every platform, with a 7-day free trial (see pricing for the current offer)
• Your credentials stored encrypted; revoke any time from the dashboard

This is the right choice for almost everyone who wants to use the tools rather than host them.

🔧 Self-hosted — free forever, your keys, your machine

Clone this repo and run it with your own platform credentials. MIT licensed, no strings, no phone-home. Your tokens never leave your machine.

The honest cost: several of the ad platforms make you apply for API access (see the table below). If you already have those credentials — you're an agency with a Google Ads developer token, a developer with a reviewed Meta app — self-hosting takes minutes. If you don't, budget days to weeks for the platforms' own approval queues, or just use the cloud.

Same code either way. The hosted product is this server plus multi-user auth, the approvals already done, and three families not in this repo yet (Merchant Center, X Ads, Bing Webmaster Tools).

Connect to your AI client

The hosted endpoint is the same for every client:

https://mcp-ads.com/mcp-slim

Sign-in is OAuth in the browser on first use; there is no API key to paste. For a self-hosted server, use the stdio config under Self-hosted quick start instead; the note under each client says where it goes.

Claude.ai and Claude Desktop. Settings → Connectors → Add custom connector, paste the URL above, then approve the sign-in window. (Guide) Self-hosted: Claude Desktop takes the command/args block in claude_desktop_config.json.

Claude Code.

claude mcp add --transport http mcp-ads https://mcp-ads.com/mcp-slim

The first tool call opens the browser sign-in. (Guide) Self-hosted: claude mcp add mcp-ads -- python /absolute/path/to/mcp-ads/server.py, or the same block in .mcp.json.

ChatGPT. Custom connectors need Developer mode:

1. Settings → Plugins → Browse plugins, and turn Developer mode on (one time).
2. Settings → Connectors → Add custom connector. Name it "MCP Ads" and paste the URL above.
3. Sign in with the Google identity that has access to your ad accounts.
4. In a new chat, open the tools menu and switch the connector on. ChatGPT keeps connectors off per chat by default.

(Guide with a recording)

Cursor. Add this to ~/.cursor/mcp.json (or .cursor/mcp.json in a project) and restart Cursor:

{  "mcpServers": {    "mcp-ads": {      "url": "https://mcp-ads.com/mcp-slim"    }  }}

(Guide) Self-hosted: the same file takes the command/args block instead of url.

Codex. Add [mcp_servers.mcp-ads] with url = "https://mcp-ads.com/mcp-slim" to ~/.codex/config.toml. (Guide)

Gemini CLI. Add an mcp-ads entry with the URL to ~/.gemini/settings.json and run /mcp to check it. (Guide)

Grok. On grok.com: + → Connectors → New Connector → Custom, paste the URL and finish the sign-in. (Guide)

Self-hosted quick start

git clone https://github.com/RoyAzran/mcp-ads.gitcd mcp-adspip install -r requirements.txtcp .env.example .env      # fill in the platforms you usepython server.py doctor   # shows what's connected

Then add it to Claude Desktop (claude_desktop_config.json) or Claude Code (.mcp.json):

{  "mcpServers": {    "mcp-ads": {      "command": "python",      "args": ["/absolute/path/to/mcp-ads/server.py"]    }  }}

That's the whole install. Ask your agent "audit my Google Ads wasted spend" and watch it go.

Docker instead:

cp .env.example .envdocker compose up -d      # HTTP server on 127.0.0.1:8000, MCP endpoint at /mcp

HTTP mode without Docker: python server.py serve --transport http

OAuth helper for the bearer-token platforms (opens a browser, stores and auto-refreshes the token):

python server.py auth snapchat_ads

What works out of the box, honestly

Self-hosting means holding your own keys to each platform. Some hand them over in minutes; some make you apply. This is the real picture:

If that table makes you tired, that's exactly what the cloud version is for — the approvals are done, you just connect your accounts.

Tool categories

One dispatcher per family. Counts are registered actions in the hosted build (September 2026); each dispatcher links to its API reference page.

Total: 3,363. A self-hosted build registers the same families except the three marked hosted only; organic social has 12 actions here rather than 32, and a few other counts differ by a handful as actions are added or hidden. list_actions always reports what your build actually has.

How the tool surface works

Your agent sees one dispatcher per platform plus list_actions:

1. list_actions(category="meta_ads_action", search="budget")2. meta_ads_action(action="<name from step 1>", params={...})

• Write actions create everything paused unless explicitly told to launch (see Safety model)
• --surface full registers every action as its own MCP tool instead — useful for programmatic clients, far too many tools for a chat client

Example prompts

• "Which search terms spent over $50 with no conversions last month? Add the worst ten as exact-match negatives."
• "Pause ad sets with frequency above 3 in my main Meta account."
• "Compare Meta and Google CPA for last week against GA4 conversions."
• "Which landing pages get paid traffic but convert below the site average?"
• "Build a paused Search campaign for 'emergency plumber' with three ad groups and a $50 daily budget."
• "What are my competitors running on Facebook and Instagram right now?"
• "Check my GTM container for paused or unused tags."
• "Write this month's executive report across Google Ads, Meta and GA4."

The NPX CLI

No Python on the client machine? The CLI is a ~300-line stdio proxy (cli/) that forwards to any MCP Ads server:

# against your self-hosted serverMCP_ADS_BASE_URL=http://127.0.0.1:8000 npx -y @marketingmcp/cli
# against the cloud (API key from mcp-ads.com/manage)MCP_ADS_API_KEY=mcp_live_xxx npx -y @marketingmcp/cli

A note on Google Ads Transparency

tools/google_ads_transparency.py reads Google's public Ads Transparency Center by calling the endpoints the website itself uses. Google documents no API for this data; the endpoints can change or start refusing without notice, and automated access to them is not something Google has blessed. The family is off by default — set GOOGLE_ADS_TRANSPARENCY_ENABLED=true only after reading that module's docstring. The Meta Ad Library tools use Meta's official API and carry no such caveat.

Safety model

• Approval before writes. Each dispatcher carries MCP tool annotations: a family with any write action is marked readOnlyHint: false, destructiveHint: true. The check fails closed, so hidden write actions count too, and paid image generation is never marked read-only. Clients such as Claude and ChatGPT use those hints to ask you before the call runs. The server's own instructions also tell the model to state a write plainly and get confirmation before it touches live spend or public content.
• Created paused. Campaign, ad set, ad group and ad creation default to PAUSED on Google Ads and Meta. Nothing starts spending until you ask for it to be enabled.
• Read/write role gating. permissions.py lists the write tools in WRITE_TOOLS, and a name heuristic fails closed for write verbs missing from that list. Write tools call require_editor(), which refuses the viewer role; only editor and admin can write. On the hosted service, a viewer seat gets that refusal.
• Running read-only when self-hosted. There is no config switch yet. The self-host build has one user, SelfHostPrincipal in credentials_env.py, whose role is admin. Change that role to "viewer" and every write tool refuses. For a second layer, use read-only credentials where the platform offers them: a Meta token with only ads_read, and the read-only GA4 and Search Console scopes (analytics.readonly, webmasters.readonly). The Google Ads API has no read-only scope, so for Google Ads the role change is the control.

Security posture

• Credentials come from your .env and an optional Fernet-encrypted local store (MCP_ADS_SECRET_KEY). Nothing phones home: telemetry is entirely opt-in and off unless you configure your own analytics project.
• HTTP mode is unauthenticated by design (it's you, on loopback). The server refuses to bind a public interface unless you set MCP_ADS_ALLOW_PUBLIC_BIND=true, because it holds credentials that can spend real money.
• CI runs a boundary gate (tests/test_no_private_imports.py) that fails on anything shaped like a secret entering the tree.

See SECURITY.md for reporting.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. The check scripts in scripts/ are the review bar: each one pins a failure that actually happened in production, and python scripts/check_tools_execute.py before a PR catches most of what review would.

One thing we can't help with: obtaining platform API access. Those approval queues belong to Google, Meta, LinkedIn, TikTok and Microsoft.

FAQ

Is there a Google Ads MCP server that can make changes? Yes, this one. Google's official Google Ads MCP server is read-only (reporting and diagnostics). MCP Ads has 755 Google Ads actions, including creating campaigns, ad groups, keywords and responsive search ads, adding negatives, and changing budgets and bids. New campaigns are created paused.

Does it work with ChatGPT? Yes. Add it as a custom connector in Developer mode with the URL https://mcp-ads.com/mcp-slim; the steps are under Connect to your AI client and in the ChatGPT guide.

Do I need a developer token? Not for the hosted version: its Google Ads developer token and the other platform approvals are already in place, and you only sign in. For self-hosted Google Ads, yes: you need your own Google Ads API developer token, and Google reviews the application (see What works out of the box).

License

MIT.