MCP vs API in 2026: Who Actually Runs Your Ad Optimization?

Fifteen years in agencies leave one impression above the others: how much of the optimization work is still done by hand. Budgets moved campaign by campaign. Placements excluded one at a time. The same checks run again every week, on every account.

Meanwhile, the experts doing it could spend that time on work with far more value: stepping back, looking at the strategy, thinking about where the business should go next.

So when Meta opened its ads MCP server this spring, the excitement was easy to understand. The question now comes up in almost every conversation with advertisers and agencies: can you plug Claude or ChatGPT into your ad accounts through MCP and let it handle everything, optimization included?

It’s a great promise. But once you look at what optimizing really involves, the answer is less simple than it sounds. This guide goes through what the official documentation of the platforms says today, and where the line sits.

Two ways to connect an AI to your ad accounts

Every ad platform has an API. It is the door software uses to read an account and change it, with its own rules and limits.

MCP sits on top of that door. An MCP server gives a language model a list of tools, and the model decides which one to use based on what you ask in the chat. Behind each tool, the server still calls the platform API.

So MCP doesn’t replace the API. It puts the model in charge of using it. The protocol says so itself: its tools are designed to be model-controlled.

That’s great for a conversation. It matters a lot more once money starts moving.

What the platforms offer today

The big platforms made very different choices, and this changes every month.

Google released its official Google Ads MCP server in October 2025. In September 2026, its documentation still describes it as strictly read-only: it can’t change bids, pause campaigns or create assets. A Google contributor to the project has said write support is planned.

Meta went further. Its ads AI connectors, in open beta since April 2026, let an assistant create and edit campaigns, ad sets and ads. Meta added sensible safeguards: new objects are created paused, activation asks for confirmation, and since July business owners can set rules on what agents are allowed to do. TikTok and Amazon Ads also have servers that can write.

So the old shortcut, MCP reads and the API writes, is already outdated. It was a convenient one, ours included, until the documentation said otherwise.

Being able to write is not the same as optimizing

A write is one action. Optimization is everything around it. In the teams that do this well, it always comes back to the same loop:

1. 1Read the accounts continuously, not only when someone asks.
2. 2Compute what is going on: pacing, wasted spend, creative fatigue, what each campaign really optimizes for.
3. 3Propose a precise change.
4. 4Let a person approve it.
5. 5Apply it cleanly.
6. 6Measure what it did, and remember it for next time.

A read without a write is a report. A write without a measurement is a bet.

Where a conversation reaches its limits

An assistant connected through MCP is good at the first step and, on some platforms, at the fifth. The trouble starts when the model is also expected to carry the rest of the loop.

What the model can see

Everything a tool returns goes into the conversation, and a conversation has a size limit. Claude Code, for example, caps what an MCP tool can bring into the conversation at 25,000 tokens by default. A search terms report on an active account over three months is thousands of rows. What doesn’t fit is either left out or summarized before anyone really looks at it.

Who does the math

In a chat, the model often calculates the CPA or the ROAS itself. Anthropic’s own engineers point out that with large data, models may be more likely to make mistakes when copying data between tool calls. In a conversation, a wrong number is annoying. In an optimization, it moves budget in the wrong direction. Ask the same question twice and you may not get the same figure either, which is hard to defend in front of a client.

What happens between two questions

An assistant answers when you talk to it. Campaigns don’t wait. A budget that starts overspending on a Saturday night or an ad that collapses mid-week will sit there until someone opens a chat and thinks to ask.

Rules and follow-up

Clicking allow on an action answers the question: can the agent do this? It doesn’t answer: should we raise this budget by 40% on a campaign still in its learning phase? Those business rules have to live somewhere and apply to every change. And once the change is live, someone has to check, a week later, whether it actually worked. None of that is part of the protocol.

What gets remembered

An API gives you today’s state. The history is thin: on Google Ads, the change history available through the API covers the last 30 days. A conversation forgets even faster, since it starts from nothing every time you open a new one. So the questions that matter most on a mature account stay unanswered: what did we change last quarter, what did it do, and which audiences have already worn out here.

Whose side the connector is on

A platform connector is built by the platform. Meta’s server, for instance, exposes an opportunity score and its own recommendations. That’s useful, and it optimizes inside Meta. Arbitrating between Google and Meta is a different exercise: it needs the same definitions on both sides, and a source of truth outside the platforms, like your analytics or your e-commerce data.

Security

A model with write access reads a lot of text it doesn’t control: campaign names, ad copy, search terms. Meta’s own guidance for its MCP servers is clear: don’t grant write or manage scopes to AI agents that process untrusted input.

MCP is still a good tool

None of this is an argument against MCP. We use these assistants every day, and the servers are useful.

• Exploring an account you don’t know yet.
• Getting a quick answer before a client call.
• Preparing the first draft of a report.
• Checking a setting without opening the ad manager.
• Setting up a paused campaign on Meta that you will review.

What changes the equation is what happens when you try to turn that into a routine. Connectors, prompts and scripts have to be maintained: API versions move, tokens expire, the platforms add tools every month. It usually ends up resting on the one person in the team who set it all up.

In two words: MCP to ask, a system to optimize. Most teams will end up using both.

How we approach it at Laier

This loop is what we built Laier around, because it is exactly what teams rebuild by hand, account after account.

Laier reads your accounts continuously. The numbers are computed by code, never generated by the model, and they keep the same definitions from one week to the next. Laier proposes precise moves, like shifting budget from one campaign to another, excluding a placement or cutting what doesn’t perform. You approve. Laier applies the move on the connected platforms, records the decision and follows what it produced.

That last part is what turns a change into an optimization. The next proposal knows what the previous one did on this account.

And if your team already works in Claude or ChatGPT, you don’t have to choose. We are bringing Laier into your assistant through MCP: the calculations and the guardrails stay in our code, the reasoning and the conversation stay where your team already is.