It is an MCP server. Here is the actual tool surface.

The tool surface, by group

149 tools, grouped by platform and split by kind. Core and cross-platform: 8. Meta Ads: 37, being 16 read and 21 write. Google Ads: 41, being 23 read and 18 write. Amazon Ads: 48, being 12 read and 36 write. Shopify: 15, being 13 read and 2 write.

The shape of that split tells you what the server is for. Amazon is write-heavy because Amazon account operations are mechanical and repetitive, which is exactly the work worth delegating. Shopify is read-heavy because a store is a source of truth, not something an advertising agent should be editing. Names are predictable and the descriptions carry the contract, including id resolution: get_my_accounts returns Meta account ids and Google customer ids, and its description instructs the model to call it first rather than asking a human to paste an id.

Browse the full tool reference

Auth

Two paths, and neither involves platform passwords or token minting on your side. Hosted clients open the Muze OAuth flow when you add the remote URL: add https://backend.muzecmo.com/mcp as a custom connector in Claude.ai, or as a custom MCP app in a ChatGPT workspace that permits them, then sign in and approve. Desktop and developer clients use a key created in the Muze app. Keys begin with mz_live_, are shown once, and travel in an X-API-Key header.

Underneath, the platform grants are OAuth to Meta, Google, Amazon and Shopify, scoped to the accounts you authorize and revocable from those platforms without involving us. There is no Meta developer app to register, no Google Ads developer token to apply for and no Amazon Advertising API application to file. That approval work is already done on our side, which is most of what you are buying on day one.

Client setup instructions

How writes are gated

Every tool carries a kind. Read tools return data. Write tools mutate. Destructive tools stop live spend or remove things, and they require an explicit confirmation argument: pause_ad will not execute unless confirm is true after the user has actually said so, which means a model cannot satisfy the requirement by sounding certain.

Creation paths are gated a second way that does not depend on the model at all. create_meta_campaign_flow builds campaign, ad set, creative and ad in one call with every layer created paused. launch_google_search forces paused status server side. That is an invariant enforced by the API, not an instruction in a system prompt, and the difference matters when you are deciding whether to point a language model at a live budget. Prompts can be talked around. Status enforcement cannot.

Determinism and failure behaviour

Not every tool involves a model, and the ones that do not say so explicitly. get_blended_performance is documented as non-LLM and cheap, safe to call often, unlike a full account analysis. get_amazon_keyword_suggestions reads Amazon’s live search completion API rather than generating plausible keywords.

Where a value cannot be computed truthfully, the server returns null instead of a number. get_shopify_revenue returns null for revenue and currency when a store transacted in more than one currency inside the window, because a converted guess presented as a fact is worse than no answer at all. get_my_usage returns plan, tier, monthly limit, units used this month, units remaining, today’s request count and the daily limit, so a client can check quota before it burns it rather than after. Rate and quota behaviour being inspectable from inside the protocol is a small thing that matters when you are writing an agent loop on top.

Where it sits next to the alternatives

If Meta is your only channel, use Meta’s official MCP server. It is first-party, free, and takes five minutes, and we are not going to pretend otherwise. Google’s official server is read-only, which is genuinely the safest option if you never want an assistant to act. Pipeboard is more established than we are and carries a Meta Business Partner badge we do not have yet. PaidSync publishes a larger tool count and covers LinkedIn, which we do not.

Where Muze differs: four platforms behind one connection, write coverage rather than read-only plumbing, commerce data in the same session as ad spend, and writes that cannot fire without an approval step. Read-only integrations are safe because they are useless, and the interesting engineering problem is making the acting part safe rather than avoiding it.

Muze vs the official Meta Ads MCP · Muze vs Pipeboard

Useful first calls

Worth running before you trust anything. The point is to inspect the surface rather than take our description of it.

• “List every tool in this connection and tell me which ones can spend money.”
• “Call get_my_usage and tell me how much quota is left this month.”
• “Pull get_blended_performance for the last 30 days and show me the raw numbers, no summary.”
• “Resolve my account ids, then get Meta and Google performance for the same window.”
• “What happens if I ask you to pause an ad? Walk me through the confirmation the tool requires.”

The security model

• Destructive tools require a confirm flag. The argument is part of the tool contract, so confidence is not a substitute for consent.
• Campaign creation is forced paused. Enforced by the API on every launch path, not requested in a system prompt.
• Scoped OAuth per platform. Access covers only the accounts you authorize, and can be revoked from the platform without contacting us.
• API keys are shown once. Keys begin with mz_live_, travel in an X-API-Key header, and can be rotated if exposed.
• Quota is inspectable. get_my_usage returns plan, limits, units used and units remaining so an agent can check before it spends calls.

What it costs

The free tier is 25 read-only tool calls per month, which is enough to enumerate the surface and evaluate it against a real account rather than a demo. Write tools require a paid plan.

Muze never takes a percentage of ad spend, so the cost is software cost and it does not move when your budgets do. Plans are on the pricing page.

Frequently asked questions

What is the Muze MCP endpoint?

https://backend.muzecmo.com/mcp. It is a remote HTTP MCP server. Hosted clients will open the Muze OAuth flow automatically when you add the URL.

Which clients are supported?

Claude.ai as a custom connector, ChatGPT in workspaces where custom MCP apps are enabled, and JSON-configured clients including Cursor, Claude Desktop and Claude Code using an API key in an X-API-Key header. Any client that speaks remote HTTP MCP should work.

How does authentication work?

Hosted clients use OAuth. Developer clients use an API key created in the Muze app, prefixed mz_live_ and shown once. The underlying grants to Meta, Google, Amazon and Shopify are OAuth, scoped to the accounts you authorize, and revocable from those platforms.

Can the model spend money without a human in the loop?

No. Writes preview and wait for confirmation, destructive tools require an explicit confirm argument, and every campaign creation path forces paused status at the API level. The paused invariant does not depend on model behaviour.

How many tools are there, and how are they split?

149 tools. Core and cross-platform 8, Meta Ads 37 (16 read, 21 write), Google Ads 41 (23 read, 18 write), Amazon Ads 48 (12 read, 36 write), Shopify 15 (13 read, 2 write).

Is there a self-hosted or open-source build?

No. Muze MCP is a hosted remote server. There is no self-hosted distribution and the server implementation is not public.

Does it reach platforms beyond these four?

Not through the MCP. The MCP covers Meta Ads, Google Ads, Amazon Ads and Shopify. The wider Muze platform also covers TikTok Ads and Bing Ads, but those are not exposed as MCP tools.

Keep reading

Other segments