---
title: "The Agentic Ecosystem Atlas: A Thanksgiving Gift to Ad Tech"
date: 2026-11-15
summary: "A free, sourced, living map of which ad platforms let AI agents in, and what an agent can read, write, optimize or activate through each exact interface. My thank-you to the ad tech community, open to people and agents alike."
standfirst: "The Agentic Ecosystem Atlas answers the question every agentic project in advertising starts with: what can an AI agent actually read, write, optimize or activate on this platform, through this exact interface, and who has to approve it. It is free, sourced claim by claim, corrected in public and open to agents through an index and an MCP server. This Thanksgiving it is a thank-you to the ad tech community, and a shared starting point for the conversation everyone is about to have."
canonical: https://nofluffadvisory.com/writing/the-agentic-ecosystem-atlas/
---

*Every week someone asks me the same question in a different costume: can an AI agent actually do this on that platform? Not "does the vendor say agentic." Can it change the budget. Can it put the campaign live. Who has to approve it. I built the answer, I put the receipts next to every cell, and this Thanksgiving I am giving it to the industry for free.*

---

## The question nobody could answer in one place

The agentic wave in advertising produced a lot of announcements and very few contracts. A press release says a platform "opens to AI agents." The developer page says the server is read-only. A partner deck says the agent "optimizes." The tool list shows a reporting query. If you are buying, building or selling, the gap between those four documents is where your project dies.

So I stopped answering the question one Slack thread at a time and built the [Agentic Ecosystem Atlas](/agentic-atlas/): a living map of which ad platforms expose interfaces to AI agents (MCP servers, REST and GraphQL APIs, SDKs, AdCP, AAMP and A2A agents, OpenRTB deals) and what an agent can do through each **exact** interface. Not the company. The interface. Meta's ads MCP server and Meta's Marketing API are two different answers, and the Atlas treats them that way.

As of today (dataset v2026-11-15.29) it covers 174 companies, 477 interfaces and 1,309 cited sources. It has shipped 43 releases since it went live on 12 November, and it keeps 125 dated public corrections.

> [figure: Previous screen]

## How 174 made the cut

The Atlas is a shortlist, not a census. Across the research behind it I looked at about 500 distinct companies: the original research kit, a market-map sweep, a search for missing MCP servers, and the vendor registers of my 14 published research papers, which on their own name 428 companies.

Of those 428, 52 were already in the Atlas and 12 were retired or absorbed brands. The other 364 went through a first screen in eight batches: S1 for the 47 highest-priority names, S2 to S5 for the next 186, and L1 to L3 for the 131 named only once. The rule was strict. Full research only when the screen found an MCP server or an AdCP, AAMP or A2A agent; a vendor with only a classic API was parked. In S1 alone, 17 of 47 qualified, 13 had only a classic API, 13 could not be settled from public pages, and 4 showed no interface at all. Across all eight batches, 88 companies made it through and brought 243 interfaces with them.

So the map is not big for the sake of it. Every company on it earned its place with something an agent can call, and the long list behind it is why I trust the short one.

## Four verbs, four states, one interface at a time

The core of the Atlas is a simple grid. For every interface it asks four questions: can an agent **read** account data, **write** changes to account objects, **optimize** (apply a bid, budget or pacing change, not just recommend one), and **activate** (make something go live: spend starts, an audience is delivered, a deal or creative is published).

Each answer is one of four states. Documented. Limited, meaning narrower or conditional. Excluded, which needs an explicit vendor rule-out, a complete published contract that lacks the capability, or a design that cannot do it. And unknown. That last one matters most. Unknown is not no. Half the bad decisions I see in agentic projects start with someone reading silence in the docs as a refusal, or reading a marketing verb as a capability.

## Start with the job you need done

Most people do not arrive with a vendor in mind. They arrive with a job. So the Atlas starts there: change a campaign, change bids or budgets, or put an audience live. Pick one and the map narrows to the interfaces that qualify, checked on the same interface and on the object it writes. A Google Ads MCP server that writes assets and labels does not qualify for "change a campaign," even though the company obviously runs campaigns. A different API at the same company does not count. An MCP server does not borrow the authority of the API behind it.

Today 90 interfaces qualify for changing a campaign, 81 for changing bids, budgets or pacing, and 34 for putting an audience live. Each of the three also has its own page: [campaign change](/agentic-atlas/tasks/campaign-change/), [bids and budgets](/agentic-atlas/tasks/bid-budget-change/), [audience activation](/agentic-atlas/tasks/audience-activation/).

From there the path is four steps, shown as one row at the top of the page: choose a task, add two to four interfaces to compare, compare and copy the vendor-call brief, then track what changed.

## Side by side, with the answer first

For the questions I hear most, I wrote the comparison down. Five [side-by-side pages](/agentic-atlas/compare/) give the short answer per task, the reason, any condition, and the published operation where the vendor names one:

- [Meta, Google, Amazon and Microsoft ads MCP servers](/agentic-atlas/compare/walled-garden-ads-mcp/)
- [TikTok, Snap, X and Meta](/agentic-atlas/compare/social-ads-mcp/)
- [Magnite, PubMatic, Equativ and Google Ad Manager via Prebid](/agentic-atlas/compare/sell-side-agents/)
- [Yahoo DSP, FreeWheel, Apostra and Vibe](/agentic-atlas/compare/dsp-buying-agents/)
- [Kevel's MCP server against the API behind it](/agentic-atlas/compare/kevel-mcp-vs-api/)

> [figure: The walled-garden comparison: one server is read-only by contract, one writes only assets and labels, two create campaigns. Every answer comes from the same engine as the map.]

The walled-garden page is a good example of why the interface level matters. Four first-party MCP servers look alike from the outside. Microsoft's is read-only by its own contract. Google's writes assets and labels, not campaigns. Meta's and Amazon's create campaigns and change budgets. If your plan assumed parity, the table tells you in ten seconds.

## Every cell shows its receipts

Open any record and you get the notes and the numbered sources behind each cell. For 44 high-demand interfaces I went further and read the vendor's pages claim by claim: each capability, the access rule, the sign-in method and the date now names the one source that decides it, with a short paraphrase and the day it was read. Where the public pages do not settle a claim, the record says so instead of guessing.

The record also tells you how to connect. Access is derived from the reviewed notes into plain tiers: open, any customer account, account plus an entitlement or enablement step, partners or by request, invite or staged rollout, or not established. Sign-in says whether the vendor names the method or the Atlas only infers it. Dates say what kind of date they are (first release, GA, beta start, announcement, a directory listing) and keep their precision, so a listing that says "September" stays September.

> [figure: Every company has a plain page for citing and sharing. Six high-demand interfaces carry a setup guide built from the vendor's own pages.]

## From answer to setup

Knowing an interface can do the job is half the work. For six of the most requested ones (Google Ads, Microsoft Advertising, Snap, PubMatic, Magnite and Amazon Ads) the company page carries a setup guide: who provides it, where it is hosted, the endpoint, sign-in, the official example config with its placeholders, a first read-only call, and whether a sandbox and pricing are published. Nothing in a guide is a guess, and none of it holds a credential.

If you are planning a whole chain rather than one interface, the Workspace has a stack worksheet. Pick a buying, selling, data and measurement interface, and it lists what each step documents, what access and approval apply, and what is still missing. It is a gap checklist, not a compatibility certificate. One step passing its own check does not make the chain work end to end.

> [figure: Workspace: your comparison, your shortlists (kept in your browser, never in a link), and the stack worksheet.]

## Learn the ecosystem, not just the vendors

Two tabs are there for people who are new to agentic advertising, or who have to explain it to a CFO. **Ecosystem** draws the roles in an agentic buy and walks through stories: an open-marketplace buy, an audience bridge, a curated deal, a closed platform, why read is not write, and how measurement feeds the next decision. Switch on authority mode and drop a real interface into a port to see its documented state replace the teaching assumption. **Learn** lays out the protocol layers and twelve jobs a buying or selling agent needs done, with each company's best state for each one.

> [figure: Learn: the protocol layers and the jobs, so a new reader can place MCP, AdCP, AAMP and OpenRTB in one picture.]

## Built to stay honest

A map of a market this fast goes stale in weeks unless it is built not to. Three things keep it current. Every interface carries three separate clocks: when a person last reviewed its claims, when the weekly watch last read its cited pages, and when its endpoint was last probed. Only a dated human review moves the first one. When a cited page changes, the record says "source changed, review pending" until someone reads it. And every change, correction and withdrawal has a stable ID in a public feed, with an RSS version.

I also asked to be checked. Two independent external audits went through the Atlas end to end, and the second came back saying six of the first audit's ten fixes passed and four were only partly fixed, plus three new bugs. All of it is fixed and documented, and the corrections it produced are public. There is a [method page](/agentic-atlas/method/) that explains how every state, task answer, badge and date is decided, and a public set of 28 [difficult cases](/agentic-atlas/qa-cases.json) (an MCP server that does not inherit its API's authority, a budget forecast that is not a budget change, a read-only contract that makes "excluded" legitimate) with the expected answer, the reason and the Atlas's current answer. The build fails if any of them disagree.

> [figure: The method page. If you disagree with a call, this is where you find out how it was made.]

## Made for agents too

It would be odd to build a map of agent interfaces that agents could not read. So the Atlas answers machines as well as people, one question at a time:

- a remote MCP server at `https://nofluffadvisory.com/mcp/atlas`, no key needed, with `find_companies`, `get_company` and `atlas_info`, typed outputs and honest errors, and a per-address limit so it serves questions, not bulk copies;
- a [sample](/agentic-atlas/atlas-sample.json) of the data format, ten companies frozen on 15 November, with its [JSON Schema](/agentic-atlas/atlas-index.schema.json);
- a [change feed](/agentic-atlas/changes.json) and an [RSS feed](/agentic-atlas/changes.xml) with stable IDs, saying what changed and when;
- and the Ask Evgeny assistant on this site, which answers Atlas questions with the interface, the source and the dataset version.

Point your agent at the MCP server and ask it which sellers it can reach over AdCP, or which interfaces can change a budget with a person approving. It gets the same answer you would.

And on top of the map sits the [Agentic Maturity Index](/agentic-atlas/maturity/): every company in the Atlas scored 0 to 100 on what its interfaces let an agent do, built only from the same documented facts. It shows who is moving fastest, how each category compares and how scores shift week to week, it is rebuilt every week, and it has a Present mode for putting it on a screen in a meeting. A [sample](/agentic-atlas/maturity.json) is open data, and the [full dataset](/agentic-atlas/maturity/#mx-data) is yours on request.

> [figure: The Agentic Maturity Index, rebuilt weekly from the Atlas data.]

## Why give it away

Ad tech has been good to me. The people in it taught me most of what I know, argued with me in working groups, and answered my questions at odd hours. Thanksgiving is a good moment to give something back, and the most useful thing I can give is a shared, sourced starting point for the conversation everyone is about to have.

So the Atlas is free. No account, no paywall, no email gate. Use it to vet a vendor before the call, to scope a build, to settle an argument, or to teach a new hire. Cite it, link to it, and point your agents at it. What I ask in return is simple: when you know something the Atlas does not, or you think a call is wrong, use the correction link on any record. A primary source changes a state. That is the deal, and it is how this stays the place people go first for anything agentic in ad tech.

Happy Thanksgiving. Start at [nofluffadvisory.com/agentic-atlas](/agentic-atlas/).

<style>
.aeag, .aeaf { margin: 2rem 0; font-family: var(--font-sans); }
.aeag-head { display: flex; flex-wrap: wrap; align-items: center; justify-content: space-between; gap: 0.6rem 1rem; margin-bottom: 0.8rem; }
.aeag-kicker { margin: 0; font-size: 0.68rem; font-weight: 700; letter-spacing: 0.14em; text-transform: uppercase; color: var(--ink-light); }
.aeag-nav { display: flex; align-items: center; gap: 0.35rem; }
.aeag-nav[hidden] { display: none; }
.aeag-btn { min-width: 2rem; height: 2rem; padding: 0 0.5rem; border: 1px solid var(--rule); border-radius: 999px; background: var(--paper); color: var(--ink); font: 600 0.8rem/1 var(--font-sans); cursor: pointer; }
.aeag-btn:hover, .aeag-btn:focus-visible { border-color: var(--accent); color: var(--accent); }
.aeag-btn:disabled { opacity: 0.35; cursor: default; }
.aeag-track { list-style: none; margin: 0; padding: 0 0 0.4rem; display: grid; grid-auto-flow: column; grid-auto-columns: 100%; gap: 1rem; overflow-x: auto; scroll-snap-type: x mandatory; scroll-behavior: smooth; overscroll-behavior-x: contain; scrollbar-width: thin; }
.aeag-slide { scroll-snap-align: start; min-width: 0; }
.aeag-shot { display: block; border: 1px solid var(--rule); border-radius: 10px; overflow: hidden; background: #fff; box-shadow: 0 1px 1px rgba(26,28,32,.06), 0 10px 24px -12px rgba(26,28,32,.35); }
.aeag-shot img { display: block; width: 100%; height: auto; }
.aeag-shot:hover, .aeag-shot:focus-visible { border-color: var(--accent); }
.aeag-cap { margin: 0.6rem 0 0; font-size: 0.88rem; line-height: 1.5; color: var(--ink); }
.aeag-count { min-width: 3rem; text-align: center; font-size: 0.8rem; font-weight: 600; color: var(--ink-muted); }
.aeag figcaption, .aeaf figcaption { margin-top: 0.6rem; font-size: 0.82rem; line-height: 1.5; color: var(--ink-muted); }
@media print {
  .aeag-nav { display: none !important; }
  .aeag-track { display: block; overflow: visible; }
  .aeag-slide { margin-bottom: 1rem; break-inside: avoid; }
}
</style>
<script is:inline>
(function () {
  function carousel(fig) {
    var track = fig.querySelector('[data-aeag-track]'), slides = track.children, count = fig.querySelector('[data-aeag-count]');
    var prev = fig.querySelector('[data-aeag-prev]'), next = fig.querySelector('[data-aeag-next]');
    function at() { return Math.round(track.scrollLeft / Math.max(1, track.clientWidth)); }
    function sync() { var k = at(); count.textContent = (k + 1) + ' / ' + slides.length; prev.disabled = k === 0; next.disabled = k === slides.length - 1; }
    function go(k) { track.scrollTo({ left: Math.max(0, Math.min(slides.length - 1, k)) * track.clientWidth }); }
    fig.querySelector('[data-aeag-nav]').hidden = false;
    prev.addEventListener('click', function () { go(at() - 1); });
    next.addEventListener('click', function () { go(at() + 1); });
    track.addEventListener('scroll', function () { window.requestAnimationFrame(sync); }, { passive: true });
    fig.addEventListener('keydown', function (e) {
      if (e.key === 'ArrowRight') { go(at() + 1); e.preventDefault(); }
      if (e.key === 'ArrowLeft') { go(at() - 1); e.preventDefault(); }
    });
    sync();
  }
  function init() { document.querySelectorAll('[data-aeag]').forEach(carousel); }
  if (document.readyState === 'loading') document.addEventListener('DOMContentLoaded', init);
  else init();
})();
</script>
