Product

Topic Engine Content Engine Code-to-Docs AEO Rank Research-Driven Content Autopilot Visibility Tracking Website & Migration Content Factory AEO Agency Pricing

Resources

Browse all resources → Latest articles Case Studies Blog FAQ Use Cases Knowledge Base Research Docs

We run our content studio over MCP: 4 lessons for your integration

Quick Answer

Connect agents to your content system through an MCP server whose tools each do one bounded job: read freely, save one piece at a time, and leave publishing to a person.

That is how our own studio works. Source code comes in through read-only tokens. Agents answer from it with no outside claims, and a deterministic check flags leaks and outside links before anything goes live. One broad update tool would undo all of that in a single careless call.

Did this answer your question?

Key Points

  • Anthropic released MCP as an open standard in November 2024; Dheeraj Sharma's 2025 post showed 50 custom integrations falling to 15 connections, though nothing in that limits what each tool may write.
  • In AEO Content's October 2026 repository pilot, the code-only rule cut web sources per article from 30 to 38 down to 0, with every product claim correct.
  • An Excella engineer on the Federal Tech Podcast estimated a 5% or 10% chance an injected request still succeeds, so publish and delete belong with a person.
Three things marketing leaders believe about MCP. Myth or fact?
Call each one, then see how other readers called it.
1 Once an agent connects over MCP, the protocol itself limits what it may write.
2 An agent can draft and save content without ever holding the power to publish.
3 A connected agent will raise your AI visibility on its own.

Watch the 1-minute version

We run our content studio over MCP: 4 lessons for your integration

The same article in 52 seconds. Watch on the AEO Content YouTube channel

A single plate waits on a kitchen pass under a warm lamp beside an open laptop, suggesting one small checked piece of work at a time

One dish to the pass at a time: the same rule we apply to every agent write over MCP.

I want to begin with the dull part, because the dull part is where the safety lives. When we connect a software product's code to our studio, the connection is read-only: GitHub through the AEO Content GitHub App or a read-only token, Bitbucket through a read-only token. The index is refreshed every night, and it reads only what changed since the last commit. Nothing on that side can be written. Of course, that is the whole idea.

The writing side is narrower still. An agent answers one knowledge-base question from the code alone, with no web research, no outside statistics and no brand claims. It describes what a user sees on screen, in the screen's own labels, and it never puts code, file names, paths or identifiers on the page. Before anything is published, a deterministic check flags any leak and any outside link.

I think of it the way I think of a good kitchen pass. The cook plates one dish. Someone looks at it, warm under the lamp, and only then does it go out to the room.

The lessons below come from running content this way, and they are written for a marketing director, founder or agency owner deciding how far to let agents into their own systems. The first is why every write tool should do exactly one bounded thing. After that come the places integrations actually fail, what an agent should be able to reach, and the one decision worth making before Monday, when the next request to give an AI agent full access lands in your inbox.

We wrote the same two knowledge-base questions twice: once with web research, and once under a rule that let the writer read only the product's code. The first round pulled 30 to 38 web sources into each article and carried 15 to 18 first-person opinions apiece. The second round had zero web sources, zero opinions and zero outside mentions. In both rounds, every product claim checked line by line against the code was correct.

So accuracy held. The noise vanished. What changed was only what the writer was allowed to touch.

That pilot is the quiet center of this article. An agent's tools decide what reaches the page, more than the model does, and more than any prompt.

If you run marketing or an agency, the decision in front of you this morning is simple to state. You can give AI agents one broad tool that updates anything, or many small tools that each change one thing and leave a record behind. The first is quicker to build. The second is the one that holds up on a bad Thursday, when nobody is watching and an agent is halfway through its queue.

Want agents that write only what you allow?

Reading a product end to end once caught an admin page promising an authenticator app while the real sign-in sent a one-time code by text and email. Bounded inputs catch that kind of drift.

AEO Content's MCP Setup and Integration maps every action your agents can take, then splits it into bounded, logged writes.

Book an MCP scoping review

What does it mean to run a content studio over MCP?

Our studio, built on 26,577 AI visibility audits, runs over MCP: agents draft and save work only through tools we define, and each tool limits what it may touch.

Before you connect anything, answer three questions in order:

  1. Which systems would an agent need to read?
  2. Which single objects would it need to write, one at a time?
  3. Which actions must a person still approve before anything goes live?

Model Context Protocol (MCP) is an open standard that lets an AI client, such as ChatGPT or Claude, call tools on a server you control. That is all it is. The protocol carries the request. The tool decides what happens next.

The volume is why we bother. Every audit runs against our AEO Rank standard of 53 criteria across 5 pillars, and we track whether ChatGPT, Claude, Perplexity and Google AI Overviews name our clients. Across our records sit 11,000+ domains scored across 15 sectors and 28 categories. Each of those becomes a reason to read something, compare it, and write one careful paragraph in response. I could not do that by hand. To be honest, nobody could.

So the agents read. They read briefs and evidence and audit findings, the way a cook reads the market before deciding on supper. Then they write, but only a little at a time. A queue built that way is easy to check: each write is one small, separate step, so you can see exactly what changed and stop a mistake before it spreads.

The infrastructure under it was designed by a full-stack engineer and content infrastructure architect with 20 years of building enterprise systems. In our experience, that background shows up less in what the agents can do and more in what they cannot. A review of 4 sources from our own records suggests the same shape everywhere: most of the work is reading, and the writing is narrow on purpose.

The common assumption is that wiring an agent to a content system is the hard part. It is not. Connection is cheap now. Upkeep is the real cost, since platforms change: in 2023 Meta replaced Facebook Creator Studio with Meta Business Suite. If your agents will draft from your own records, our page on research-driven content built on evidence walks through the standard we use to build articles from evidence rather than from memory.

For a marketing director, the decision is plain. You are not choosing whether agents can reach your CMS. They can. You are choosing how small each door should be.

And that question has a history, because the protocol was built to fix one problem and quietly left another one on the table.

Why was MCP created, and what problem did it leave open?

MCP was built to cut how many connections you maintain, not to limit what each connection may write. That gap matters when our promise is being named in AI answers in 100 days.

The original problem was arithmetic. In a 2025 Substack post, Dheeraj Sharma laid it out plainly: before MCP, 5 AI apps and 10 data sources meant 50 custom integrations, and under MCP the same setup needed 15 connections. He called the old way an "integration tax," and he was right about the bill. Every update to a tool could break every assistant wired to it.

So the protocol fixed the count. It did not fix the scope.

Sharma also wrote that most developers spent 3 to 5 weeks in trial and error before their first working integration. That number tells you something useful. The hard weeks were not spent on the handshake. They were spent deciding what each tool should be allowed to do, and finding out, usually late, what happened when it was allowed too much.

His post said, of tools under MCP, "nothing runs without approval." I like the sentence. I do not quite trust it. Approval lives in the application that hosts the agent, not in the protocol, and a host can ask once and never ask again. A single permission granted on a Tuesday morning can quietly cover every write that follows it.

Fewer connections do not mean fewer risks. Each remaining connection can carry more weight.

Alex and I bring a combined 40+ years of SEO and content infrastructure experience to this work, and that long view makes me suspicious of any fix that only moves the problem one room over. We have seen integrations before. The shape of the failure rarely changes. Only the address does.

Think about what a client actually buys. When Imagine Dental Arts saw AI-referred new patients rise 900%, from about 5 a month to 50, the thing that mattered to them was the page a patient read, not the pipe that carried it there. A protocol never wrote a sentence anyone wanted to read. Our guarantee, that a client is named in AI answers in 100 days or we work free until they are, depends on that page being right every time. One careless overwrite, made by an agent with too wide a tool, can undo weeks of careful work in a second.

For a buyer, the practical reading is short. Count connections if you like. Then count what each one can change. Our AEO knowledge base guides cover how those pages are built; the question here is who is allowed to touch them, and that answer belongs, oddly enough, to whoever builds the server.

Lesson 1: Why should every write tool do exactly one bounded thing?

Whoever builds the MCP server decides its shape. Give each write tool one bounded action, such as saving a single block, and nothing wider.

On a 2025 episode of the Tech Ease podcast, the host broke an MCP setup into four parts: the LLM, the MCP client, the MCP server, and the tool itself. Then came the line that matters most for a marketing team. The companies that make the tools, the host said, are responsible for building their own MCP servers. The episode also called MCP, at the time, a blueprint still being finalized, and said these integrations needed custom technical work rather than an app store download.

Read that twice. The protocol hands you the pen. It does not tell you how wide to make the page.

I founded LiveHelpNow, which was recognized on the Inc. 5000 at number 84, and I have built enough customer software to distrust any button labeled "update everything." In our studio, an agent saves one block of an article at a time. If it needs to change that block, it rewrites that block, and only that block. The block is the unit of work, and I would like to keep it that way.

Why so strict? Because the pages are worth something. With Understood Care, a client, Google AI Overview citations rose 500%, from 22 to 132, in about five months. Gains like that are built one paragraph at a time. They can be lost the same way, only faster.

Volume makes the case sharper. Our plans run from Growth at $99 a month for 10 articles to Business Plus at $1,000 a month for 44 articles, with Premium at $250 and Business at $500 in between. Every one of those articles is many small writes. Multiply a broad tool by that many calls and the odds stop being theoretical.

A narrow tool fails small. A broad tool fails everywhere at once.

That is the whole lesson, really, though it took me longer to believe than to say. When a bounded write goes wrong, you see one bad block, you fix it, and the rest of the article sits there untouched, like plates still warm on the table after one glass has tipped over. When a broad update goes wrong, you are reading the entire piece again, line by line, trying to find what moved. If you are weighing this for your own stack, our answers to common AEO questions cover how the pieces fit together.

This week, find the broadest write tool your agents can call. Write down every field it can change. Then split it until each piece changes one thing, and leave the reading tools wide open, because reading is where agents earn their keep and where the next lesson begins.

Where do MCP integrations actually fail, and what should an agent be able to reach?

A manager's bonus shows up in a chat. An agent deletes a help desk ticket. A tool description changes after someone approved it. A broken connection causes none of these failures.

An AI engineer at Excella gave the plainest example on the Federal Tech Podcast in October 2025: a naive SharePoint integration without role-based access controls. An assistant meant to see the bonus policy could also surface a manager's actual bonus from last year. The engineer also cited agents deleting help desk tickets or changing records in a system of record, and called responsible permission work on MCP servers hard and time-consuming.

The protocol won't do that work for you. Dheeraj Sharma, who teaches an MCP course, wrote that under MCP "nothing runs without approval." The specification is softer than that. It says there "SHOULD always be a human in the loop with the ability to deny tool invocations," and adds that "the protocol itself does not mandate any specific user interaction model." Whether a person sees a prompt first depends on the host.

Even where approval exists, it happens once. OWASP's MCP Top 10 lists tool poisoning among its top risks. CData, which sells a managed MCP gateway, describes the attack: "An attacker buries instructions inside a tool's description or response, and the model follows them as though a developer wrote them." CData adds that "almost no one goes back to check whether the description quietly changed." Data the agent should never have seen also hides well:

"The record looks the same as any other, so it ends up in the answer."

CData, on off-limits data reaching a model, 2026

From the outside, CData says, such failures "end up looking like hallucination." A permissions problem can pass for a model problem.

Prevention has limits too. The Excella engineer estimated that the best current defenses still leave roughly a 5% or 10% chance an injected request succeeds. The engineer pointed to the "lethal trifecta" coined in June 2025 by Simon Willison, an AI researcher and Django co-creator: an agent combining access to private data, exposure to untrusted content and the ability to communicate externally. "Failing to understand this can let an attacker steal your data," Willison wrote. If some injections get through, the useful question is what the agent can reach when one does.

What went wrongWhere it comes fromScope control to consider
Overshared files reach an answerSharePoint example, Federal Tech PodcastRole-based access controls on what the server can read
Records deleted or changedFederal Tech PodcastRead-only access when the job is reading. One bounded write when it is not.
Injected instructions succeed5% or 10% estimate, Federal Tech PodcastRemove one of the three conditions in Willison's trifecta
Tool description changes after approvalOWASP MCP Top 10, as reported by CDataRecheck descriptions whenever a server updates

Our studio gave us one test of narrow scope. In October 2026 we connected it through read-only tokens to two Bitbucket repositories behind a live-chat software product, indexing 1,368 text files and deliberately excluding 367 backup and retired template files. We wrote the same two knowledge-base answers twice: first with web research, then under a rule limiting the writer to the code.

Web sources per article, with web research 30 to 38
Web sources per article, code-only rule 0
First-person opinions per article, with web research 15 to 18
First-person opinions per article, code-only rule 0
What narrower input removed. The same two knowledge-base questions written twice in our repository pilot with a live-chat software company, October 2026.

Accuracy held. In both rounds every product claim checked line by line against the code was correct, and no source code, file name or identifier reached the page. What changed was everything riding along with the answer: with the web available, about a third of one article was regulation, attack statistics and industry framing. Under the code-only rule, outside mentions also fell to zero, and a deterministic check flagged any leaked identifier or outside link before publishing.

Narrow input also made the work easier to audit. The study step tied each proposed topic to one to four specific files and line ranges, and dropped a topic when the controls it would describe were not yet wired up in the code. After a day of commits, the re-index re-read only the 14 files that changed. Our pilot tested writing, with no injection or tool poisoning attempts, so it cannot say how often those attacks succeed.

All of these failures happen in the same place: the gap between what a task needs and what a server lets an agent reach. MCP standardizes the wiring and leaves that gap for each builder and host to close. When we narrowed it, the agent brought in less, and everything it brought in could be traced to a file.

  • Write down each tool's job and remove read access to anything it does not need. If the job is reading, issue a read-only token.
  • Check whether the agent holds private data, reads untrusted content and can send data out. If it does all three, remove one.
  • Ask the vendor of your MCP host which tool calls prompt a person before they run, and whether anyone can switch those prompts off.
  • Re-review tool descriptions after every server update, and treat an unexplained change as a security event.
  • Put a deterministic check between the agent's draft and the publish button that flags outside links and leaked identifiers.

How we checked this

Our sources were the MCP specification, a federal technology podcast interview with an AI engineer, Simon Willison's original post, an article from an MCP gateway vendor and an MCP explainer. The pilot figures are our own, from October 2026. The podcast evidence comes from a machine transcript, so we paraphrase it instead of quoting it, and the 5% or 10% injection figure is one engineer's estimate. We report the OWASP listing as CData describes it. CData sells a managed MCP gateway, Sharma sells MCP training, and we sell content produced through our MCP studio, so each of us has a stake. Our pilot covered one company and two questions and did not test attacks. Still unknown: how often tool poisoning happens in practice, and whether narrower scope lowers injection success rates.

  1. Federal Tech Podcast, episode 276 on Model Context Protocol and federal software integration, October 23, 2025.
  2. Model Context Protocol, specification page on tools, version 2025-06-18.
  3. Simon Willison, the lethal trifecta for AI agents, June 16, 2025.
  4. Dibyendu Datta, CData, article on choosing an enterprise MCP server, October 13, 2026.
  5. Dheeraj Sharma, Substack post on why MCP exists, October 27, 2025.
  6. AEO Content, repository knowledge-base pilot with a live-chat software company, October 2026.

What should you decide about MCP before Monday?

Decide that every agent write will be small, single-purpose and logged, and that publishing stays with a person. That choice costs some design time and protects every page you already own.

Anthropic released MCP as an open standard in November 2024 to end the integration tax. It did. What it handed back, quietly, was the question of scope, and scope is now yours to answer.

When we took the web away from our writer, the opinions fell to nothing and the facts stayed true. Narrow inputs did that. Narrow tools do the same for writes. In our experience, the safest piece of our own setup is also the dullest: repositories connect through read-only tokens, and a deterministic check flags any leaked code or outside link before a page is published.

I think the next year or two will make this ordinary. Buyers will ask vendors what each server can write and what it logs, not whether a server exists.

Start with the broadest tool you have. It is usually the one nobody remembers approving.

Summarize This Article With AI

Open this article in your preferred AI engine for an instant summary.

Frequently Asked Questions

What else do people ask about connecting content systems to AI agents with MCP?

Most questions come down to three things: how to connect, what an agent may change, and whether the connection itself earns citations. It does not; the pages do.

How do I connect our content system to AI agents with MCP?

You put an MCP server in front of the system, and that server offers tools and resources to a client such as ChatGPT, Claude or Copilot. On a 2025 episode of the Federal Tech Podcast, an AI engineer from Excella described switching one on as fairly plug-and-play, much like enabling a plugin. The plugging in is easy. Deciding which tools exist is the real work.

Is it safe to give an agent write access to a CMS?

Only in small, logged pieces. The same episode noted that the specification had been out for less than a year and servers with security flaws were already appearing. It also said that holding a successful injected request to a 5% or 10% chance was roughly the best defense on offer at the time. I would not bet a live site on those odds.

What is the lethal trifecta?

The lethal trifecta is Simon Willison's term for an AI system that holds private data, reads untrusted outside content, and can send data out. A content agent that browses the web and can also publish has all three. Publishing is a way of speaking outward. I take that one away first.

Should an agent ever publish or delete content?

Not on its own. The episode warned that loose integrations could let agents delete help desk tickets or change records in a system of record. A content library is a system of record too, and a cold one to rebuild. Keep publish and delete behind a person.

Does an MCP connection improve AI citations by itself?

No. Citations follow the pages. With Seamless Chex, a client, the brand was linked in 28 of 40 buyer questions after 164 articles, and weekly AI bot visits rose from 883 to 4,945. The connection only carried those articles; what the engines quoted was the writing.

How do I get help scoping our MCP setup?

Our MCP Setup and Integration service starts with the tool list: what agents may read, what they may save, and what stays with your team. Reach us through the contact page at aeocontent.ai/contact. Bring the list of systems you want connected, even a rough one scribbled on a napkin.

Written by

Michael Kansky

Co-Founder, AEO Content

Michael Kansky is a serial founder and operator and co-founder of AEO Content, where he shapes product and go-to-market strategy for an AI-search content optimization platform.

Connect on LinkedIn

Related posts

A sprawling family tree scene drawn on a large sheet spread across a wooden kitchen table, with red string pinned between some of the branching names and several small portrait frames lying loose and unconnected beside

JSON-LD that validates can still leave your entity graph broken

Five laptops on a dark desk, each showing a different audit dashboard for the same printed web page

Same URL, five AEO audit tools, five grades: what to disclose

Two laptops comparing an old website with its rebuilt version next to a printed list of URLs being checked off

How to keep your rankings when you rebuild your website with AI

Pricing

Simple, flat monthly pricing.

Everything included. No per-seat games. Cancel anytime: your content, your repo.

Growth

$99 /mo

Start showing up in AI engines.

Start with Growth

What's included

  • AEO Website + Cloudflare CDN
  • 10 AEO articles / month
  • 5 prompts tracked weekly
  • Monthly re-audits + alerts
  • Reddit reply drafts + mention monitoring
Most chosen

Premium

$250 /mo

The package marketing teams settle on.

Start with Premium

Everything in Growth, plus

  • 20 AEO articles / month
  • Publish to 2+ site sections
  • 20 prompts tracked daily + threat analysis
  • Bi-weekly re-audits
  • Brand voice profile + strategy call

Business

$500 /mo

Hand us your domain. We run AEO end-to-end.

Talk to us

Everything in Premium, plus

  • 30 AEO articles / month
  • Social media on 6 networks
  • Up to 100 competitors + API
  • Weekly re-audits
  • Dedicated AEO strategist