Tempio

GoMarble versus Meta Ads MCP

One agent you install in your terminal, one server you point a client at. What each project publishes about itself, side by side.

GoMarble and Meta Ads MCP, row by row
GoMarbleMeta Ads MCP
What it is"An AI marketing agent in your terminal", built by GoMarble: a plugin you install, backed by the GoMarble MCP server."A Model Context Protocol (MCP) server that lets AI assistants run your Meta Ads end to end", published by Pipeboard.
What it reachesNine sources in its README: Google Ads (Search, Shopping, PMax), Meta, TikTok, LinkedIn, Bing Ads, GA4, Shopify, Klaviyo and Search Console.Meta only. The README names it as one node of a five-server family covering Meta, Google, TikTok, Snap and Reddit, with 230+ tools combined.
What you installA plugin and a skill pack: "42 expert-authored skills" and "8 morning-workflow slash commands", added through a marketplace or with npx marketing-agent.Nothing, on the hosted path: a remote MCP URL in your client's config. A local install is documented and the README does not recommend it.
Tool surfaceThe README counts skills rather than tools; the live data arrives through the GoMarble MCP server and the count is not published in the repository.42 tools by the README's own count, 29 of them documented by name: campaigns, ad sets, ads, creatives, image upload, insights, interest, behaviour, demographic and geographic targeting, and budget schedules.
What it changesAudits, creative-fatigue detection and campaign launches. Its slash commands are described as read-only analysis with no mutations.Create and update campaigns, ad sets and ads, build and edit creatives, upload images and schedule budgets, alongside the read tools.
Sign-in"OAuth, one-click sign-in" through the GoMarble MCP server.Interactive OAuth against the hosted endpoint, or a token on the URL. Self-hosting needs your own Meta developer app.
LicenceMIT.Business Source License 1.1, which the README says "Becomes fully open source (Apache 2.0) on January 1, 2029". Running it as a competing hosted service is restricted.
Guardrails as published"Built-in guardrails — no synthetic data, no fabricated keywords, proper attribution discipline". Its eight morning commands "never execute mutations"; the create flow "builds the propose payload, asks you to confirm before posting"."Explicit confirmation on every write; new campaigns start paused". The README says the same contract holds across all five servers: "writes are explicit, new campaigns start paused where the platform supports it, and confirmation prompts look identical".

A plugin and a server are not the same purchase

This is the decision under the comparison, and it is easy to miss because both projects install from a terminal and both end up as tools in the same chat.

GoMarble's marketing agent is a plugin: skills, slash commands and a house method for reading an account, with live data arriving through GoMarble's own MCP server behind a one-click sign-in. You are adopting somebody's opinion about how a performance marketer works, and the data connection comes along with it.

Pipeboard's Meta Ads MCP is the opposite shape. It is a tool surface and nothing else — 42 tools by its own count over one platform's API, no method, no opinion, no skills. What your agent does with them is yours to define.

So the question is not which is more capable. It is whether you want an opinion or an interface.

Breadth against depth on one platform

GoMarble's README names nine sources, and four of them are not ad platforms at all: analytics, a store, an email tool and a search console. That is a deliberate shape. An audit that can see the store and the analytics alongside the ad account can say something about the funnel rather than only about the campaign.

The Meta Ads MCP goes the other way: one platform, 42 tools by its own count, down to interest and behaviour search, geographic targeting and budget schedules. If the work is inside one ad account and the detail matters — building ad sets, validating interests, scheduling budgets — that depth is the thing you are buying, and the four sibling servers cover the other platforms the same way.

A rough rule: if your question starts "why did revenue move", breadth answers it. If it starts "build me this", depth does.

The licence is the part people skip

GoMarble's marketing agent is MIT licensed: use it, fork it, ship it, no conditions worth reading twice.

The Meta Ads MCP is under the Business Source License 1.1, and its README states that it "Becomes fully open source (Apache 2.0) on January 1, 2029". In the meantime, use for individual and business purposes is free and running it as a competing hosted service is not. For most readers that distinction never comes up. For an agency thinking about wrapping it and selling access, it is the first thing to read rather than the last.

What neither repository will tell you

Neither README publishes a price, a credit allowance, a seat count or a rate limit. Both point at a website for the account and the billing, and this page carries nothing from either of those sites, because they could not be opened from the machine this page was written on.

So treat every row above as a statement about the software each project publishes, not about what either one costs to run at your volume. That is a real gap in this page and it is better stated than papered over: check both pricing pages yourself before you choose, and note the date you read them.

Read the repository, not the landing page

Everything above comes from two places anyone can open: the README of GoMarble's marketing-agent repository and the README of Pipeboard's meta-ads-mcp repository, both read on 12 September 2026.

That is a narrower source than a comparison usually draws on, and it is worth saying why it is also a better one. A README is the page a project's own engineers maintain, it changes when the software changes, and it lists the tools by name. A landing page is written once and re-read by nobody. When the two disagree, the README is the one that has to work.

It also means this page is missing things. Pricing is the obvious one. Support terms, uptime, and how either project behaves on an account with two thousand ads are others. Those are questions for a trial, not for a comparison table, and a table that pretended to answer them would be the kind of page this one is written against.

What "end to end" means on the Meta Ads MCP

The phrase in Pipeboard's README is that the server "lets AI assistants run your Meta Ads end to end", and the tool list is what makes that concrete rather than promotional: get accounts and account info, list and create campaigns, list and create ad sets, update ad sets including frequency caps, list, create and update ads, build and update creatives with single or multiple headlines and descriptions, upload images, pull insights with breakdowns and attribution windows, search interests, behaviours, demographics and geographies, list account pages, and create budget schedules.

Read as a list, that is most of a campaign build. A model with those tools and a competent operator behind it can go from a brief to a built ad set without a browser open. Whether that is what you want on a Tuesday afternoon is the real question, and it is the same question a keyboard shortcut poses: speed is only an improvement when the thing being sped up was correct.

What forty-two skills buy you

GoMarble's count is of skills, not tools, and the distinction matters. A tool is a verb the model can call. A skill is a written method for using verbs: what to look at first, what a bad number looks like, when creative fatigue is the explanation and when it is an excuse.

The README also names eight morning-workflow slash commands and describes them as read-only analysis with no mutations, which is a specific and unusually honest thing to publish. A daily audit that cannot change anything is a daily audit you can run without reading it first.

The guardrail line — "no synthetic data, no fabricated keywords, proper attribution discipline" — is the other half of the same idea. Anyone who has watched a model produce a confident keyword list for a campaign that does not exist knows why that sentence is in the README.

How to choose without a trial

Three questions settle it faster than a table.

Where is the money? If it is concentrated on one platform, the depth of a single-platform server is worth more than the breadth of an agent that touches nine sources shallowly. If it is spread, the opposite.

What is the recurring job? "Tell me what changed and why" is an audit, and audits want the analytics and the store in the same context. "Build this and launch it" is a construction job, and construction wants precise verbs.

Who is holding the keyboard? An operator who already knows what they want writes better prompts against a bare tool surface than they get from somebody else's method. Someone newer to paid media gets more from the method than from the verbs.

Sources

Every row and every quotation above comes from one of these pages, read on 12 September 2026.

Neither company's website could be opened from the machine this page was written on, so nothing that lives only there — pricing above all — appears above.

Both stop in the same place, and it is not at the write. Each README documents a confirmation before anything is posted: the Meta Ads MCP promises "explicit confirmation on every write" and starts new campaigns paused where the platform supports it, and the marketing agent builds a propose payload and asks you to confirm before posting, with its eight morning commands never executing mutations at all. What neither describes is what becomes of that decision afterwards. The confirmation is answered in the conversation and the conversation ends, so the record of who cleared what, on which account, is whatever the ad platform's own change history kept, and there is no scope to hand somebody that lets them read an account without being able to approve against it. Neither publishes a planning layer either: both are good at answering a question and acting on an account, and neither claims to hold the quarter's budget shape and reconcile it against what was actually spent.

Take the marketing agent if you want an opinionated method across a funnel: several platforms plus analytics and the store, a skill pack written by people who audit accounts for a living, and a one-click sign-in that gets you to a first audit in an afternoon. Take the Meta Ads MCP if most of your money is on one platform and you want a precise interface rather than a method: 42 tools, your own prompts, and four sibling servers when you need another platform on the same terms.

If what you need is the record

Every change is written as a draft and held at the gate. Publishing to live accounts is in private beta. Governed autonomy — approved by you, or by a rule you wrote. In private beta.

See how the gate works

Questions

Put frontier AI to work on every ad account you run.

Tempio connects your ad accounts into one workspace and hands your own AI the numbers and the tools. Every change it writes is a draft until you approve it, or until a rule you wrote does.

Request early access
  • Read and draft by default. Approve is a separate permission you grant.