---
title: What is MCP
description: What the AdCrunch MCP server lets an AI agent do with your ad accounts, and the six jobs it does.
---

**MCP** (Model Context Protocol) is an open standard. It lets an AI client, such as Claude, ChatGPT or Cursor, call tools on a server that another company runs. AdCrunch runs an MCP server. When you connect your AI client to it, the client can read your ad accounts and, on Meta, change them.

You do not learn a tool or a query language. You ask a question in your own words, for example _"Which of my Meta campaigns had the best ROAS last week?"_. The agent calls the correct AdCrunch tools, reads the answers, and explains them to you. You can also ask for a change, for example _"Pause every ad set with a CPA above 40 euros."_ The agent makes the change on Meta, and AdCrunch records it.

## The six jobs

Each job has a guide. A guide walks the job from start to end, with example prompts, and states which guardrails apply.

**[Read your accounts](/mcp/tools/read-your-accounts)**

Ask about your advertisers, campaigns, ad sets, ads and their performance,
on Meta, TikTok and Google Ads.

**[Change what runs on Meta](/mcp/tools/change-what-runs-on-meta)**

Create a campaign, an ad set, a creative or an ad, and change a status or a
budget. Each create arrives paused.

**[Plan a campaign](/mcp/tools/plan-a-campaign)**

Write down what you intend to run, on which channels, for how much, and
approve the plan before anything is created.

**[Describe a brand](/mcp/tools/describe-a-brand)**

Tell the agent who your brand is, how it sounds, who it talks to, and which
files belong to it.

**[Bring your own creative](/mcp/tools/bring-your-own-creative)**

Upload your images and videos, then put them in the library of an ad
account.

**[Reuse a playbook](/mcp/tools/reuse-a-playbook)**

Save a routine as a playbook, and run it again with one pick from a menu.

## How it works

AdCrunch hosts the server at `https://mcp.adcrunch.dev/mcp`. You install nothing. You add this address to your AI client, and you sign in to AdCrunch once. The client then calls AdCrunch tools for you, for one organization.

- **The client sees every tool.** Each client gets the same list of 51 tools. The [scopes](/mcp/auth#scopes) of a token decide which tools can run. A tool that your client has no scope for fails with the code `forbidden`, and the agent tells you so.
- **The agent never holds a provider password.** Your client holds an AdCrunch token. It never holds a Meta, TikTok or Google token.
- **Each change is on the record.** AdCrunch writes each change to a ledger before it runs. The **Activity** page of the console shows who asked, what changed, and how it ended.

[What to expect](/mcp/what-to-expect) states the rules that each tool follows. [Auth & scopes](/mcp/auth) states what an agent can change and what it cannot.

## What's next

- [Set up Claude](/mcp/setup/claude-desktop), the most common client. Other clients are in the same group.
- [What to expect](/mcp/what-to-expect): data freshness, money, queued changes, and edits that collide.
- [Prompts](/mcp/prompts): run a playbook or insert a brand from the slash menu of your client.
- [Errors](/mcp/errors): the shape of a failed tool call.