Ads MCP Server: Run and Analyse Campaigns from Your AI Assistant

An ads MCP server lets the AI assistant you already use read your ad accounts, answer questions about performance and, where you allow it, act on campaigns, all in plain language. Instead of exporting a CSV from Meta, another from Google and a third from TikTok, you ask "which accounts lost money this week?" and the assistant calls the server for the numbers. This guide explains what MCP is, what an ads MCP server can and cannot do, the guardrails that matter, how to choose one, the setup steps and where a command-line tool fits better.

It sits next to two articles we have already published. If you want the news angle on Meta's own connector, read what Meta's AI Connectors and MCP mean for tool safety. If you want every way of linking AI to ad accounts ranked from riskiest to safest, read five ways to connect AI to your ad accounts, ranked by risk. This piece is the practical guide to using an ads MCP server day to day.

What MCP is, in plain words

The Model Context Protocol is an open standard for connecting AI applications to outside systems. The official MCP introduction compares it to a USB-C port for AI applications: one standard plug, so any compatible assistant can connect to any compatible data source or tool without a custom integration each time.

Three pieces are involved:

• The client is the AI assistant or app you type into. It has to support MCP.
• The server is the service on the other side. It publishes a list of tools, each with a name, a description and the inputs it accepts.
• The tools are the actions the server offers, such as "list ad accounts" or "get campaign insights for last week".

When you ask a question, the assistant reads the list of tools, picks the ones that fit, calls them and turns the result into an answer. According to the MCP tools specification, tools are "model-controlled", which means the model decides which tool to call based on your prompt. That is exactly why the guardrails later in this guide matter.

What an ads MCP server lets your assistant do

For advertisers, an ads MCP server turns the assistant into a front end for your ad accounts. The useful jobs fall into three groups.

Read performance across accounts

The first job is reading. Ask for spend, ROAS, conversions or status, and the assistant fetches them through the server. With a single-network server you get one platform. With a multi-platform server you can ask "compare Meta and TikTok, this week against last" and get both in one reply, which is the part that usually eats time when you manage ads across multiple platforms.

Answer questions you would otherwise build a report for

The second job is analysis. Instead of building a pivot table, you ask follow-up questions in sequence: which campaigns carry most of the spend, which ad sets fell below your target ROAS, which accounts need reconnecting. The assistant keeps the context of the conversation, so each question narrows the last one.

Propose changes

The third job is acting, and this is where servers differ the most. Some are read-only by design. Others offer write tools, such as pausing an ad set or adjusting a budget. A sensible pattern is that the assistant proposes a plan ("pause 4 ad sets under 1.5 ROAS, leave budgets untouched") and nothing changes until you confirm it.

Meta ads MCP and Google Ads MCP: single-network options

Both of the largest ad networks now have their own route for AI assistants.

• Meta ads MCP. Meta launched official AI Connectors and an MCP server for its ads surface on 29 April 2026, in open beta. Our breakdown of the Meta launch covers what it does, what it leaves open and why it matters for tool safety.
• Google Ads MCP. Google's developer documentation describes a Google Ads MCP server with tools for account discovery and performance reporting. The page states that "this implementation is strictly read-only", so it cannot change bids, pause campaigns or create assets. It authenticates with OAuth 2.0 or a service account, and support runs through community channels such as GitHub issues.

These are good options if you buy on one network. The limit is scope: each covers a single platform, so a team running Meta, Google, TikTok and native together ends up with several connections and still has to reconcile the numbers by hand.

Risks of connecting AI to ad accounts

Connecting an assistant to accounts that spend real money brings risks worth naming plainly.

• Unofficial access. A server that scrapes the ads interface or drives a browser session is a very different thing from one that calls the official API. Our risk ladder for AI connections explains why the connection method matters more than the AI itself.
• Over-broad credentials. A key that can do everything on every account is a key that can do damage if it leaks or if the assistant misreads a prompt.
• Unconfirmed writes. Because the model picks the tools, a vague request could trigger a change you did not intend.
• Hidden tool calls. If you cannot see which tools were called with which inputs, you cannot review what happened.

None of these is a reason to avoid MCP. They are reasons to choose the server, and the settings, with care.

Guardrails to insist on

The MCP specification itself sets the baseline. On the tools page, it says there should always be "a human in the loop with the ability to deny tool invocations", and that clients should prompt for user confirmation on sensitive operations, show tool inputs before calling the server and log tool usage for audit purposes. Servers must validate inputs, apply access controls and rate limit tool calls.

In practice, that translates into five checks:

1. Official APIs only. The server should reach each ad network through its official API, with your accounts connected by the method the network sanctions.
2. Scoped API keys. You should be able to limit a key to specific areas and actions, and grant read-only access where reading is all you need.
3. Revocation. You should be able to revoke a key at any time, and use a separate key per person, tool or environment so one revocation does not break everything.
4. Confirmation before changes. Your AI client should ask before each action. Treat any setup that changes campaigns without asking as a red flag.
5. An audit trail. Keep the client's record of tool calls, and check what the server or platform logs on its side, so you can trace who asked for what.

If you already use Wavo, the AI copilot inside Wevion, the same principle applies there: it proposes and you approve. Our guide to choosing a mode for Wavo, the AI copilot explains how approval-first works across its three modes.

How to choose an ads MCP server

Use these questions to compare options:

If you are also weighing AI tools that sit inside an ad platform rather than in your own assistant, our roundup of AI media-buyer copilots compares autonomous and approval-first options.

Where the Wevion MCP server fits

The Wevion MCP server and CLI bring Meta, Google, TikTok, Taboola, Snapchat and Outbrain into one connection. Wevion syncs data from these platforms about every 15 minutes, so the assistant reads the same numbers you see in cross-channel analytics. The MCP server and CLI are available from the Pro plan.

According to the Wevion MCP page, the server exposes campaigns, ad accounts and insights as tools. You can ask for the status and metrics of your campaigns, query accounts, insights and reports in plain language, and have the assistant act on the Wevion API within your key's permissions. The page also notes that your MCP client asks before every action.

On the trust side, the page states that:

• each API key is scoped to specific areas and actions, and can be read-only or read plus write
• a key carries its own team context and can be revoked at any time
• the server only exposes the tools your key allows
• keys are encrypted and isolated per team, and Wevion does not train models on your campaigns

A sensible way to start is a read-only key for analysis, and a separate read and write key only once you are comfortable with how the assistant proposes changes.

Setup at a high level

The exact screens depend on your AI client, but the Wevion setup follows three steps.

1. Create an API key. In the API keys section of Wevion, select the team you want the assistant to work with, then generate a key. The key carries that team. It is shown once, so store it somewhere safe.
2. Add the remote server. In your MCP client, add a remote MCP server using the Streamable HTTP (POST) transport, and send your key in the x-api-key header. Only POST and Streamable HTTP are supported, not SSE.
3. Test with a read-only question. Ask something harmless first, such as "list my ad accounts and their status", and check that the answer matches what you see in Wevion.

Two practical notes. Some AI clients need a paid plan of their own before they let you add remote connectors. And a sign-in path through AI client connector directories, where you approve scopes without handling a key, is described on the Wevion page as not open yet.

The ads CLI: for scripts and automation

A conversation is great for exploring. It is less suited to jobs you want to run the same way every morning. That is where an ads CLI comes in.

The Wevion CLI drives the Wevion API from your terminal, your scripts or a CI pipeline. Based on the MCP and CLI page:

• Install: it needs Node.js 22 or later, and installs with npm install -g @wevion/cli.
• Authenticate: run wevion login and paste an API key, or set the WEVION_API_KEY environment variable for headless runs in CI.
• Explore: wevion list shows the API endpoints grouped by area, and built-in help shows the flags of any command.
• Pipe the output: results come back as JSON, so you can pass them to tools such as jq or into your own reporting.
• Work with AI agents: wevion agent prints a primer on authentication, commands and exit codes, so a coding agent can learn the API quickly.

The CLI uses the same API keys and permissions model as the MCP server. A good habit is one dedicated key per environment, so a key used in a CI job can be revoked without touching the one your team uses in chat.

When to use which:

• MCP server for ad hoc questions, exploring a drop in performance or drafting a change with your assistant.
• CLI for scheduled checks, bulk operations you repeat, and anything that should run without a person typing.

Key takeaways

• An ads MCP server lets MCP-compatible AI assistants read, analyse and, where allowed, act on your ad accounts in plain language.
• Single-network options exist: Meta has an official MCP server, and Google documents a read-only Google Ads MCP server.
• The MCP specification recommends a human in the loop, confirmation for sensitive operations and logging of tool calls.
• Choose a server on official APIs, scoped and revocable keys, read-only defaults and the platforms it covers.
• A CLI on the same keys handles the repeatable jobs a conversation is not built for.

Try it on your own accounts

Connect your platforms, create a read-only key and ask your assistant the question you usually build a report for. Wevion offers a 14-day free trial on the Pro plan with no card needed, so you can test the MCP server and CLI on your own data. Start your free trial.

Sources (checked 22 October 2026): What is the Model Context Protocol, MCP specification: tools, Google Ads API: MCP server, Wevion MCP and CLI.