What is MCP

How the Model Context Protocol lets AI agents query AdCrunch in natural language.

MCP (Model Context Protocol) is an open standard that lets AI agents — Claude Desktop, Claude Code, Cursor, ChatGPT, and others — call structured tools exposed by external servers. AdCrunch ships an MCP server that exposes your campaigns, ads, and insights as queryable tools.

In plain terms: once you wire AdCrunch’s MCP server into your AI client, you can ask it things like “what were my top three Meta campaigns by ROAS last week?” and it knows how to call the right AdCrunch tool, pass the right parameters, and explain the answer.

What you can do with it

Ask in natural language

Skip the dashboard. Ask your AI client and get the same answer.

Compose across platforms

Compare Meta and TikTok side-by-side without exporting CSVs.

Drill into one ad group

Find the underperforming creative inside a campaign.

Plug into your workflow

Use AdCrunch from wherever your AI lives — Claude, Cursor, ChatGPT.

How it works

AdCrunch hosts a remote MCP server at https://mcp.adcrunch.dev/mcp. You authorize your AI client once via OAuth; from then on, the client calls AdCrunch’s tools on your behalf, scoped to your organization.

The tools available today

Read your accounts and performance — list_advertisers, list_entities, get_entity, query_insights; and manage reusable ad-ops playbooks — skill_list, skill_get, skill_create, skill_update, skill_delete. See the Tool reference for each.

Your playbooks also show up as prompts. Clients that support MCP prompts list every playbook in your organization — usually behind a / slash menu — so you can pick one by name and run it, instead of describing the task and hoping the assistant reaches for the right playbook. Selecting one inserts the playbook’s instructions as your message.

The same menu is built into the Adgent in the console, where picking a playbook leaves a short /handle in the message box. Add anything you want to narrow it — /weekly-report for the Acme account — and send: the assistant receives the playbook’s full instructions followed by your note.

Prompts are yours to choose: the assistant can’t invoke one on its own. It can still find and load playbooks by itself mid-task via skill_list / skill_get — the menu just adds a deliberate way for you to start.

What’s next