Hosted X MCP for AI Agents

The framing that makes this launch more than a routine developer release is that X is not shipping a feature for developers - it is repositioning itself inside the workflows that are starting to replace the feed. Gennaro Cuofano of FourWeekMBA calls it a distribution play, arguing that X's hosted MCP server repositions the platform inside every AI agent workflow rather than serving as a convenience for API users ^[2]. The logic is that distribution is migrating from the feed to the agent - if people increasingly ask an assistant "what is happening right now" instead of scrolling a timeline, then the platform that answers that question has to live inside the assistant, not on a website someone has to open.

Seen that way, the MCP server is X reclassifying its core asset. A social network sells attention against a feed. An intelligence layer sells real-time data access to whatever software needs it. By exposing post search, user lookup, and trends through a standard endpoint, X is offering itself as the real-time substrate for agents that need to know what the world is talking about this minute - a category where its live-firehose is genuinely differentiated. The strategic tell is that this required almost no new product surface: X monetizes agent access through the same usage-based API pricing it already runs, so the intelligence-layer repositioning arrives with minimal operational risk ^[2].

Technically, an MCP server is a standard adapter that lets an AI model call an external tool without the model's author writing bespoke integration code. Before this, a developer who wanted an agent to read live X data had to build their own MCP bridge, host it somewhere, and handle X's authentication and rate limits themselves. X's hosted version removes all three of those chores ^[3]. The bridge runs locally through the npm launcher (npx) with no separate install step, and credentials never leave the user's machine beyond a Bearer token, with two auth routes: app-only Bearer tokens for read-only access and OAuth 2.0 PKCE for user-authenticated actions. In practice you point a compatible client - Claude, Cursor, Grok Build, VS Code - at api.x.com/mcp and the agent can immediately search posts, read timelines and mentions, look up users, manage bookmarks, and pull trends and news.

The critical constraint is that this is read-only. The MCP tool is not compatible with X's Write API endpoints, so an agent cannot post through it ^[1]. That is not an oversight - it is the design. Reads are cheap and hard to abuse at scale; writes are where spam lives, and X prices posting deliberately high (cited around 1.5 cents per post and 20 cents per post with a link) to deter it ^[3]. Keeping the hosted MCP read-only lets X hand agents frictionless access to its data without handing them a spam cannon, which is exactly why any workflow that needs to publish still has to go through the paid, rate-limited Write API separately.

The timing is not arbitrary. GitHub, Slack, Notion, Stripe, and Salesforce already ship official MCP servers, and once a cohort of major platforms is reachable through the same protocol, being absent from it starts to cost you ^[1]. Cuofano's sharpest line captures the stakes: MCP is becoming the address book of the internet for AI agents, and presence in the index is the new SEO ^[2]. The analogy is precise - in the search era, a business that was not indexed by Google was effectively invisible to anyone looking for it. In the agent era, a platform whose data an assistant cannot reach through MCP simply does not surface when a user asks their assistant a question that data could answer.

That MCP itself is now neutral infrastructure sharpens the pressure. Anthropic introduced the protocol and then donated it to the Agentic AI Foundation, a Linux Foundation fund co-founded with Block and OpenAI, so no single vendor controls it ^[4]. A platform adopting MCP is therefore not betting on one company's ecosystem - it is joining a standard that the largest AI labs have collectively endorsed. For X, shipping a server is less an opportunity than a defensive requirement: the cost of being the one major real-time data source that agents cannot natively reach would compound quietly, one unanswered query at a time.

There is a history angle here that makes the pricing choice read as strategy rather than accident. In February 2023, X shut down its free API tiers with roughly seven days' notice and moved to $100/month Basic and $5,000/month Pro plans, a change that drove an exodus of the indie hackers and third-party app builders who had powered its developer ecosystem ^[5]. Those tiers were priced for companies, not weekend builders, and the community that left was precisely the one that experiments, prototypes, and seeds new use cases.

By 2025, X had begun testing pay-per-use pricing and shifted its developer platform to usage-based billing ^[6]. The hosted MCP lands on top of that shift, and X frames it as pay-per-use with no monthly subscription fees - a barrier low enough for individual developers and small teams to try it ^[7]. Read against the 2023 history, the arc is a quiet reversal: the same platform that priced small developers out three years ago is now offering them zero-setup, no-subscription access to its live data, metered by usage rather than gated behind a four-figure monthly floor. The read-only scope keeps that access cheap and safe to extend broadly, which is what makes re-courting the long tail of builders economically viable this time around.