How it works

AdsGPT reaches your AI agent in two forms. Both talk to the same server and act with the same account.

	MCP server	Plugin
What it is	A remote MCP server at https://mcp.adsgpt.dev	The MCP server plus the AdsGPT skills, packaged for each client
Where it works	Any client that supports remote MCP servers with OAuth	Claude, Claude Code, Codex and Grok
What the model gets	The tools	The tools and the skills

The MCP server

The server exposes the AdsGPT tools: list your accounts, report on and query them, change budgets and statuses, and plan and launch campaigns. It runs in the cloud, so nothing runs on your machine.

It offers the tools of the platforms your organization connected: the google_ads_* tools appear only when Google Ads is connected.

The plugin and its skills

Skills are instructions the model reads when a task needs them: how to plan a campaign step by step, how to write ads within each format’s limits, how to review spend, and the rules of each platform and campaign type. With them, the model asks the right questions and gets campaigns right at the first attempt.

The plugin contains only the skills, which are text, the address of the server and a manifest for each client. It has no scripts, hooks or binaries.

Signing in

Your agent signs in to AdsGPT with OAuth 2.1, the standard MCP clients use. There are no API keys to copy:

1. The first time the agent calls AdsGPT, the server answers that it needs authorization.
2. The agent opens AdsGPT’s sign-in page in your browser.
3. You sign in, choose the organization the agent works on and approve its access.
4. The agent keeps the token and refreshes it on its own.

The agent first asks for ads:read, which lets it read. The first time a tool changes something, it asks for ads:write as well. You can revoke an agent’s access from AdsGPT at any time.

Approvals

Every AdsGPT tool declares whether it only reads or changes something, and your agent uses that to decide when to ask you:

• Reads, such as reports and queries, run without asking in most clients.
• Changes, such as a budget, a status or a launch, always ask for your approval first. Your agent says what it is about to change, with the current and new values.

A new campaign is a plan, saved with save_blueprint and checked with each platform. You approve the whole plan once, with launch_blueprint, and its campaigns launch paused unless you ask otherwise.

Note

AdsGPT acts with your role in the organization. A tool fails if your role can’t do what it asks, whatever your agent tries.

What AdsGPT receives

The server receives the calls your agent makes to its tools, with their inputs, such as an account, a query, a campaign plan or ad texts. It never receives your conversation. Security and privacy has the details.