← Back to feed

Apple shipped a Model Context Protocol server for Safari this week, arriving in Safari Technology Preview 247. X — Elon Musk’s social network — launched its own hosted MCP server three days earlier on June 30. The week adds two of the most recognisable consumer platforms to a list that already includes GitHub, Slack, Notion, Stripe, and Salesforce.

MCP is an open standard introduced by Anthropic in late 2024. It defines a common interface for AI models to connect to external tools, data, and services. Before the hosted servers existed, a developer who wanted Claude or Cursor to access X’s API had to build a custom MCP server, host it, wire it to the X API, and handle authentication. That infrastructure layer is now provided by X itself. The same shift is happening in each platform that ships an official server.

What Apple’s Safari MCP Server Does

The Safari MCP server connects an agent directly to a live Safari browser window, giving it capabilities that were previously only available through manual screenshots or manual prompting:

• DOM inspection: The agent can read the current page structure without the developer taking a screenshot
• Network request access: Live traffic visible to the agent for debugging and performance analysis
• Console output: JavaScript errors and logs stream to the agent directly
• JavaScript evaluation: The agent can run code in the page context to pull performance metrics, check accessibility attributes, or verify form state

The primary use case is debugging. Apple’s framing: a developer currently cycles between browser, console, and agent prompts. With the Safari MCP server, the agent sees what the browser sees and can troubleshoot autonomously. Accessibility checking, Safari-specific compatibility testing, and performance profiling are listed as additional capabilities.

Any MCP-compatible client works — Claude, Cursor, or anything else using the protocol. Apple has not announced a paid tier or access restriction.

What X Built

X shipped two MCP servers on June 30. The first connects AI tools to the core X API using a user’s own account permissions. The second provides AI coding assistants with direct access to X’s developer documentation, covering API specs and integration guides.

The core API server covers more than 200 X API endpoints. Capabilities include searching public posts, reading timelines, looking up user profiles, analysing conversations, and monitoring trends — everything available through the existing API. Write access is explicitly excluded: the server does not expose X’s Write API endpoints, so agents cannot post to X autonomously.

Platforms with official MCP servers now include: GitHub, Slack, Notion, Stripe, Salesforce, X, and Apple. That list covers version control, team communication, knowledge management, payments infrastructure, enterprise CRM, the largest social network, and the dominant consumer browser.

What Changes

The practical shift is integration friction. Connecting Claude to the Stripe API required a custom implementation. With Stripe’s official MCP server, it requires none. Multiply that across every platform in the list and the total integration overhead for a development team drops substantially.

The more structural change is that MCP is now a deployment target for large organisations, not just a developer experiment. X and Apple shipping official servers signals that the question is no longer whether MCP will be adopted — it is which major platforms have not shipped their server yet.

For agents that operate across multiple tools and platforms simultaneously — the pattern that defines production AI deployments — MCP’s value compounds with each additional server that comes online. A single MCP-compatible agent can now reach X’s API, inspect a Safari browser window, query Stripe payment data, and reference Notion documentation without custom connectors for any of them.

Anthropic introduced the standard. The competitive dynamic that gives it staying power is that the more platforms ship official servers, the more valuable MCP compatibility becomes for the AI clients that implement it — and the harder it becomes for any individual platform to defect.