Meta's ads MCP server arrives in Databricks Marketplace, linked to Genie

From handoffs to a single conversation

Databricks describes a familiar problem for enterprise advertisers. Data teams build churn models, lifetime value (LTV) scores and attribution data, while media teams adjust budgets using numbers in Meta Ads Manager. The insights often reach campaigns only through a CSV export, a custom Marketing API integration that someone has to maintain, or a recurring weekly sync.

MCP, or Model Context Protocol, is an open standard that lets AI agents interact with tools and data sources. According to Databricks, the ads MCP server from Meta exposes 25+ tools across the campaign lifecycle. These cover creating campaigns and ad sets, changing budgets and bids, defining audiences, managing creative and catalogs, pulling performance data, and diagnosing signal quality.

Marketers already work in Genie One, where they ask questions and get answers grounded in enterprise context, including metrics and definitions curated by data teams. Adding Meta's server brings campaign reporting and advertising tools into that same conversation. The release follows the launch of Meta Conversions API on Databricks Marketplace, which helps advertisers send first-party conversion signals from Databricks to Meta.

What marketers can do with it

Databricks gives several examples. In Genie One, a marketer could prompt: "Use our churn-risk score to identify high-value customers at risk of leaving, and launch a retention campaign with an $8,000 daily budget." Genie One then searches for the right data, finds relevant context across connected tools and systems, and creates the campaign.

• Budget allocation: gross margin per order sits in Databricks, while spend and delivery metrics sit at Meta. An agent with access to both can recommend shifting budget toward ad sets that drive profitable revenue.
• Signal diagnostics: for advertisers already sending conversions through the Conversions API, agents can access event volumes, Event Match Quality and data freshness, helping answer whether purchase events dropped because sales slowed or because a pipeline changed.
• Weekly reporting: a question about which campaigns underperformed last week, and how that compares with internal revenue, becomes one prompt instead of two tools stitched together in a slide.

Governance and controls

The obvious concern is an agent taking costly or wrong actions in a production ad account. Databricks says administrators control access to the Meta connection through Unity Catalog, while Unity Gateway governs tool calls and records usage and audit logs. Enterprise data remains subject to existing Databricks permissions.

On the Meta side, advertisers set rules in Business Settings, for example blocking any budget increase over 20% or disallowing campaign creation on a given account. According to Databricks, the server checks every tool call against those rules before it executes. A call that breaks a rule is blocked, and a structured error tells the agent why. Databricks says this is enforced server-side, so it holds regardless of which model is used. The same rules can be managed through the Marketing API, which Databricks notes matters for teams running hundreds of ad accounts.

Getting started and what it means

To begin, users find the ads MCP server from Meta in Databricks Marketplace and configure a connection using a Meta user access token and the required permissions. Databricks suggests giving the agent access to relevant marketing data and starting with a question the team already asks every week.

Stephen Orban, SVP Product Partnerships and Ecosystem at Databricks, said customers want "a more direct way to put their data and AI investments to work in the decisions that drive growth." Databricks also points to CustomerLake, an Agentic CDP built natively in Databricks, as part of the same focus on customer intelligence.

For teams that already have models in Databricks and a media team working from platform reporting, Databricks positions the integration as the shortest path between having the data and the campaign reflecting it.