Checked October 30-31, 2026. I read Meta's docs and the tool vendors' pages, and connected the official server to read the tool descriptions it gives the assistant.
The short version: Meta's official Ads MCP is a solid choice if you run one or a few Facebook and Instagram accounts and want to ask about results and make simple changes from Claude or ChatGPT. It falls short once you run many clients, need Google Ads in the same conversation, or want a before/after preview, a change log with undo and memory of what each client agreed to. That's the gap AdSparky fills, and for agencies on Meta + Google it's what I'd use. Free beta, no card: adsparky.com.
AdSparky is our product - I'm the founder, so every claim about another tool links to its source, and I say where AdSparky falls short.
What the official Meta Ads MCP is#
MCP (Model Context Protocol) is the standard that lets an AI assistant use outside tools. Meta runs its own MCP server at mcp.facebook.com/ads and calls the whole thing "Meta ads AI connectors". A connector is just a service plugged into your assistant: add the URL, sign in with Facebook, and the assistant sees your ad accounts.
Timeline:
- April 29, 2026: Meta announced the MCP server and a CLI as an open beta, supporting Claude and ChatGPT (per ppc.land).
- July 16, 2026: Meta opened the server to any developer with their own Meta app and described rules for AI agents (Meta for Developers blog, updated September 22, 2026).
- August: Jon Loomer covered new security settings in Business Settings, under Integrations, "Ads MCP Server".
What it can do#
The docs group tools into seven areas:
- Reporting and performance.
- Creating and managing campaigns, ad sets and ads.
- Product catalogs and data sources.
- Signals and datasets (e.g. Conversions API data quality - events sent from your server).
- Help and troubleshooting.
- A/B tests and lift studies (measuring how many sales the ads actually caused).
- Account activity log.
Jon Loomer counted "nearly 100" tools. Meta says it keeps adding tools and rolls them out gradually, so your set may differ from a colleague's.
How safety works#
There are several layers, and they're worth separating.
Drafts. In the tool descriptions the server gave me after connecting, draft mode was on. New campaigns land as drafts in Ads Manager and go live only when published, and they're created paused. Exception: pausing a live ad applies right away.
Deleting. The edit tool accepts a DELETED status. Its own description says that's permanent and that the assistant should ask the user first. So deletion is possible, and what guards it is an instruction to the model, not a hard block.
MCP server rules. Per ad account you can block the agent from creating campaigns, ad sets and ads, changing targeting, creative or status, or all of it. You can also cap budget changes by percentage, amount or ceiling (the docs' example blocks raises over 20%). The docs note the rules are in "limited availability", and businesses outside the program get an error.
Client-side permissions. In Claude you can set each tool to "always allow", "needs approval" or "block". That works with any MCP server, not just Meta's.
What I didn't find in the docs: a before/after preview of a change, a log of AI changes with a reason and one-step undo, or a default budget guard (you set the rules yourself, where available).
Where it works and what it costs#
Meta lists Claude and ChatGPT, and the docs say it works with "any MCP-compatible AI agent". In Claude you add it as a custom connector; per Anthropic's help center, custom connectors are on all plans, with Free limited to one. For ChatGPT, Meta's docs describe connecting through developer mode with your own Meta app ID. Agencies managing other businesses' data through their own app need app review and advanced access to the ads_mcp_management permission.
I found no pricing or fee information on Meta's pages.
Official Meta MCP vs AdSparky vs Pipeboard#
| Official Meta Ads MCP | AdSparky | Pipeboard | |
|---|---|---|---|
| Platforms | Meta only | ✓ Meta + Google Ads (+ GA4, GTM, Merchant Center) | Meta, Google, TikTok, Snap, Reddit and more |
| Writes | ✓ | ✓ | ✓ |
| Before a change | Draft in Ads Manager | ✓ Dry run + before/after preview + human approval | Declares confirmation on every write |
| Deleting | Possible (DELETED status) | ✗ Never, pauses only | No delete tool found in README |
| Guardrails | Server rules (limited availability) | ✓ Per-client limits, AI can only tighten; max +30% / -50% budget per operation | Tokens scoped to accounts and tools (paid plans) |
| History and undo | Meta account activity log | ✓ Log with author and reason, undo usually one step | Audit log export on Enterprise |
| Many clients | Account list | ✓ Morning review across all clients | Account slots by plan |
| Client memory | Not found in docs | ✓ Saved rules, what counts as a lead | Not found |
| Data location | Meta | ✓ EU (Ireland, Frankfurt) | Not checked |
| Price | No pricing found | Free beta | Free plan + paid tiers |
Full list of eight servers: best Meta Ads MCP server. Detailed comparison with Pipeboard: Pipeboard alternative.
See it on your own clients: AdSparky free beta.
When the official MCP is enough#
- You run your own business or one or two accounts.
- You advertise only on Facebook and Instagram.
- You mostly want to ask about results, creatives and pixel issues.
- You have a product catalog and want to diagnose it in chat - the official server is genuinely broad here, broader than AdSparky.
- You're comfortable setting write tools to "needs approval" in Claude and checking drafts before publishing.
It's a good tool from the source, with no third party holding access to your account.
When to choose AdSparky#
- You run a dozen clients, not one. Each morning you see where things are burning across every account, urgent first, instead of clicking through Ads Manager one client at a time. This is the part I use every single day.
- Your clients also run Google Ads. Compare cost per lead across Meta and Google in one sentence. The official server is Meta only.
- You want approval built into the tool. Every change is a dry run, then a before/after preview, then a human approves - in chat or in the panel queue. Not a setting you have to remember to flip on each machine.
- Deleting should not be an option. AdSparky only pauses. The official server can set DELETED.
- You need guardrails today, for every client. Per-client budget caps and untouchable campaigns, which the AI can only tighten. Meta's server rules are in limited availability.
- You answer to clients. Who changed the budget, when and why - and one step to roll it back. Plus PDF reports.
- The assistant should know the client. "No ads to students", "only phone calls count as leads". Saved once, applied in every conversation.
- Juniors prepare, seniors approve. Team roles plus an approval queue. Automations only prepare changes unless the owner enables autopilot.
- GDPR matters to your clients. Engine and database in the EU, encrypted keys, agency data separated, no training on your data.
- It's a two-minute setup. Paste one URL into Claude or ChatGPT, sign in, start working.
To be fair: AdSparky is a beta, it covers Google Ads and Meta Ads only, and Meta's own server also covers catalogs, lift studies and signal diagnostics.
FAQ#
Is the official Meta Ads MCP safe for client accounts?#
It uses drafts, creates new campaigns paused, and its rules can block budget changes and creation. But the rules are in limited availability and deleting is technically possible. On client accounts, set write tools to "needs approval" and start with read-only access.
Can I limit which accounts the assistant sees?#
When connecting in Claude you can pick specific Pages, Instagram accounts and business portfolios instead of everything. Jon Loomer points out that Business Settings shows every account exposed to connectors, and you can change allowed actions per account.
Does the official MCP remember client agreements?#
I didn't find client memory in the docs. The assistant remembers what fits in the conversation; in Claude you can add notes to a project.
Is AdSparky better than the official Meta MCP for agencies?#
For multi-client agency work, yes - that's what it's built for: Google next to Meta, a review across all clients, per-client limits, client memory, approvals and a change log with undo. If you run one Meta account and need deep catalog or lift-study tools, the official server covers that.
Can I use the official MCP and AdSparky together?#
Technically yes. On client accounts I wouldn't: a change made through one server won't appear in the other's history, and the assistant may pick the wrong tool.
Sources#
All checked October 30-31, 2026.
- Meta for Developers, Ads MCP Server overview: https://developers.facebook.com/documentation/ads-commerce/ads-ai-connectors/ads-mcp-server/ads-mcp-server-overview
- Meta for Developers, getting started: https://developers.facebook.com/documentation/ads-commerce/ads-ai-connectors/ads-mcp-server/ads-mcp-server-get-started
- Meta for Developers, MCP server rules: https://developers.facebook.com/documentation/ads-commerce/ads-ai-connectors/ads-mcp-server/ads-mcp-server-rules-best-practices
- Meta for Developers blog, July 16, 2026: https://developers.facebook.com/blog/post/2026/07/16/meta-ads-mcp-server/
- Official server tool descriptions (visible after connecting): https://mcp.facebook.com/ads
- ppc.land on the April 29, 2026 launch: https://ppc.land/meta-opens-its-ad-system-to-claude-and-chatgpt-with-new-ai-connectors/
- Jon Loomer, connector security settings: https://www.jonloomer.com/meta-ads-ai-connectors-security-controls/
- Anthropic, custom connectors: https://support.claude.com/en/articles/11175166-getting-started-with-custom-connectors-using-remote-mcp
- Pipeboard, pricing and README: https://pipeboard.co/pricing and https://github.com/pipeboard-co/meta-ads-mcp
- AdSparky: https://www.adsparky.com/
Running Meta and Google for many clients? Try AdSparky on your own accounts - free beta, no card: adsparky.com.