For Agents to get work done, they first need to be connected to the world.

On June 30, X (formerly Twitter) officially launched its Hosted MCP Server, enabling all MCP-compatible AI tools -- Claude, Cursor, Grok Build, etc. -- to directly access X platform data through the standard protocol. Searching tweets, reading posts, querying user info, analyzing trends -- capabilities that previously required developers to set up their own MCP servers, integrate APIs, and handle authentication -- are now available out of the box.

This isn't a feature update; it's a signal that a standard is being established.

MCP: From Open-Source Project to Mainstream Standard

MCP (Model Context Protocol) was originally open-sourced by Anthropic in late 2024, aiming to define a universal standard for AI models to connect with external tools and services. Simply put: previously, every AI tool needed separate integration code to connect to GitHub, Slack, or Notion; with MCP, one adaptation connects to all MCP-supported services.

In just over a year, MCP adoption has exceeded expectations. GitHub launched an official MCP server, Slack integrated MCP support, Notion open-sourced its MCP server, and Stripe and Salesforce launched official endpoints. According to previous reports, MCP monthly downloads reached 97 million. X's entry means one of the world's most important real-time information networks has officially opened its doors to Agents.

X emphasized two points in its announcement: first, Hosted MCP doesn't support the Write API -- Agents can only read, not post -- which blocks the path to automated spam; second, MCP won't bypass X's API rules, and spam behavior will still be restricted. Earlier this year, X raised API posting prices ($0.015 per regular post, $0.20 per post with links), using economic leverage to curb automated abuse.

The Infrastructure for an "Agent Internet" Is Taking Shape

The deeper significance of X launching MCP is that it accelerates the formation of an "Agent Internet."

Today's internet was designed for humans -- browsers, web pages, apps, buttons are all for people to click. But AI Agents need a different kind of internet: structured data, standardized interfaces, services callable by programs. MCP is becoming the HTTP protocol of this "Agent Internet" -- a universal connectivity layer that lets different Agents and different services talk to each other.

Think about it: when Cursor can search X for technical discussions via MCP to get the latest dev insights, when Claude can check GitHub code issues via MCP, when enterprise Agents can read Salesforce data, initiate Stripe payments, and send Slack messages via MCP -- at that point Agents are no longer "assistants in a chat window" but genuine digital employees capable of executing tasks across platforms.

This also explains why Amazon announced $1 billion for FDE the same day -- when Agents' connectivity infrastructure (MCP) and cost infrastructure (affordable models like Sonnet 5) are both maturing, the last mile of enterprise Agent deployment is: someone to help you assemble these pieces and run your business workflows.

Of course, MCP's普及 also brings new problems: permission management, data security, Agent identity authentication, cross-platform operation auditing -- these questions don't have standard answers yet. Research reported by Ars Technica the same day showed AI browsers can be deceived by "dream attacks" -- telling an LLM that 2+2=5 is enough to bypass guardrails. When Agents can connect to more real services, security challenges only intensify.

But the direction is clear: MCP is becoming the de facto standard for AI tool interconnection, and X's entry is a major step toward that standard going mainstream. Looking back, July 1, 2026 may well be remembered as the key day the Agent Internet moved from concept to infrastructure.

See you tomorrow.