Meta has launched the Ads MCP Server, wiring AI assistants directly into advertising accounts. Connected agents can not only analyse campaigns but also create, edit and delete them. The service runs at the endpoint mcp.facebook.com/ads and works with any assistant that speaks the Model Context Protocol.
Seven tool groups cover everyday campaign work
The server bundles seven groups of tools. Alongside creating and editing campaigns, ad sets and individual ads, it allows custom audiences to be maintained, product catalogs to be built and faults in product data to be fixed. For reporting, metrics such as spend, impressions, click-through rate and return on ad spend are available, broken down by date range, audience attributes and platform.
Three further groups address day-to-day operations. A signal diagnostic reports the health and quality of transmitted measurement data and is meant to help prioritise investment. A search function reaches into the articles of the Meta Business Help Center without users leaving the assistant. Finally, A/B tests and conversion lift studies can be created and retrieved, and activity logs for the ad account inspected. Meta says further tools will follow gradually. Until now, developers had to build such workflows themselves against the Marketing API; the server instead exposes them as ready-made, clearly scoped commands. The Model Context Protocol serves as the vendor-neutral standard through which assistants address third-party services.
Agencies need advanced access
Access is open in principle to anyone running a Meta app in the developer dashboard, where the matching use case has to be added. Two sign-in routes are available: automatic token handling via OAuth with Facebook Login for Business, or a token obtained beforehand. The second route targets developers who have already solved authentication elsewhere.
Requirements are stricter for agencies and partners managing accounts on behalf of other businesses. Per Meta’s documentation, they need advanced access to the ads_mcp_management permission and must pass an app review to obtain it. Meta thereby draws a line between reaching into one’s own ad accounts and administering someone else’s budget. No separate price is charged for the server; costs accrue through the ad bookings themselves and through the token consumption of the assistant in use. The announcement is credited to Meta staff Sydney Levitan and Sanjay Patel. The company gives no fixed date for completing the rollout of additional tools.
Partners report figures in seconds, gaps remain
Meta names several partners already running the server. Snowflake links its own customer and sales data to campaign performance, Triple Whale operates an assistant for e-commerce that validates changes and surfaces anomalies, and Salesforce automates campaign operations through its Agentforce agents. The vendor AdManage cites concrete timings: in an account holding 270 ads, a search took 2.4 seconds, reviewing performance trends 1.3 seconds, and the assistant delivered five recommendations in under five seconds. Those figures come from the vendor itself and are independently unverified.
Gaps remain. According to the marketing service provider Admove, access does not extend into Meta’s optimisation layer: bidding logic, automatic audience expansion and the Advantage+ system stay out of reach. Local files cannot be attached, so creative assets must be hosted publicly. Token consumption also weighs in, because the assistant loads the descriptions of every tool at the start of a session. For recurring routine work, Admove therefore recommends scripted routes over the conversational format.
What matters now is how tightly companies fence in that write access. An assistant able to create and delete campaigns on its own reaches straight into live budgets – a layer of control Meta leaves to its customers. How well the permission tiers hold up in daily use should become clear once the first agencies start managing third-party accounts under advanced access.
