X's New Ads MCP Server Gives AI Agents Write Access to Your Campaigns

X's advertising API now has a front door built for AI agents instead of developers. On August 21, X's business team confirmed the X Ads MCP server is live at ads-api.x.com/mcp, a hosted endpoint that lets Claude, Grok, Claude Code, or any other MCP-compatible client read and write campaign data using plain-language requests instead of custom API integration code.

This isn't X's first MCP server. The company shipped a general hosted MCP on June 30, exposing more than 200 endpoints across the whole platform, posting, search, timelines, direct messages, to any agent with a user's permissions. The Ads MCP is narrower and newer: a dedicated surface for the advertising API specifically, aimed at people who manage campaign spend rather than people who post.

Twenty-three tools, four jobs

X's own documentation lays out the server's tools in four groups. Account and read operations cover ads accounts, campaigns, line items, funding instruments, promoted posts, and targeting criteria, the baseline data an agent needs before it can say anything useful. Analytics tools, get_account_stats and get_campaign_reach, let an agent "pull performance stats and reach estimates conversationally," which is the part most advertisers will actually use first: asking an agent to summarize how a campaign performed instead of exporting a report and reading it yourself. Targeting search tools look up interests and geographic locations by name. Write operations handle campaign and line item creation, activation controls, targeting changes, and promoted post creation, the part that turns this from a reporting tool into something that can actually spend money on your behalf.

Why every new campaign starts paused

The detail worth noting if you're the one managing spend: X's documentation is explicit that "campaigns and line items are always created PAUSED." An agent can build a full campaign structure, targeting, budget, creative assignment, but nothing goes live until a human flips it on. That's a deliberate checkpoint, not an oversight, and it's the same instinct that shows up across every serious agentic-commerce integration this year, give the agent enough access to be useful and a hard stop before it can do something irreversible.

Authentication is scoped the same way. Advertisers register an X Developer app and connect via OAuth 2.0, with separate ads.read, ads.write, and offline.access scopes. The agent runs on the user's own token, so it only ever sees the ad accounts that specific user's login can already reach. There's no shared service account sitting behind the MCP server with broader access than the person using it.

Every ad platform is racing through the same door

X isn't moving first here. Meta shipped an official ads MCP server with 29 tools in April. Google has run a read-only version since October 2025. TikTok announced its own in May, and Snapchat and Pinterest have shipped comparable integrations in recent months. What's notable isn't that X joined the list, it's what every platform on it is quietly conceding at the same time: none of them is trying to keep advertisers working inside their own dashboard anymore. The strategy has shifted from owning the interface to staying the system of record underneath whatever AI tool the advertiser is already running the actual strategy through. It's the kind of business-function integration that the protocol's own move toward simpler, more scalable servers was built to support at this scale, dashboards becoming optional, the API becoming the product.

What it means if you're building or buying with agents

If you manage ad spend on X, connecting an agent through this MCP server is a genuinely low-risk way to start, the paused-by-default behavior means the worst case is a campaign that sits inert until you approve it, not one that quietly starts spending. The setup itself is documented for Grok Build, Grok's web client, and Claude Code specifically, with a manual token-minting path via curl for anything else.

If you're building agent tooling more broadly, X's Ads MCP is a clean reference for a pattern worth copying in your own integrations: read access that's genuinely useful on its own, write access that's real but defaults to a safe state, and auth that inherits the calling user's actual permissions instead of a broader service credential. That's the shape a well-designed MCP server for anything with financial or operational consequences should probably take, whether it's ad spend, infrastructure changes, or anything else you'd want a human to glance at before it goes live.

The bigger pattern here is one this site has tracked since Agent Plugins standardized how agent tooling gets packaged: the infrastructure for handing agents real operational access, not just chat, is being built out function by function, platform by platform, mostly with the same safety defaults each time. Ads happened to be next.

Sources: X Ads MCP documentation, Social Media Today: X launches MCP server, TechCrunch: X now offers an MCP server to make its platform easier for AI tools to use.