X shipped an MCP server — and it deliberately can't post
X launched a hosted MCP server that lets Claude, Cursor, and Grok read the platform through your account — but not write. Why that read-only line is the right call for AI integrations.
On June 30, X quietly shipped a hosted MCP server. The interesting detail isn't that it exists — GitHub, Slack, Stripe, Notion, and Salesforce already have one. It's the line X drew: the server can read, but it can't post.
What actually happened
X launched a hosted Model Context Protocol server (docs at docs.x.com/tools/mcp) that lets MCP-compatible tools — Claude, Cursor, Grok Build — talk to the X API using your own account's permissions. Search posts, look up users, analyze trends: all available. Writing is not. X explicitly excluded its Write API, so an agent connected through the server cannot post autonomously — or at all.
Meanwhile the protocol underneath is hardening. The MCP spec's 2026-07-28 revision — the largest since launch — drops session state from the protocol core (any request can hit any server instance, so you can run behind a plain round-robin load balancer), formalizes an extensions framework, and tightens authorization toward OAuth 2.0 and OpenID Connect, including mandatory issuer validation. MCP is turning into boring, load-balanced plumbing. That's a compliment.
Why it matters for your business
The read-only boundary is the whole lesson. Hooking an AI tool into a live business system is easy to demo and easy to regret. The failure mode isn't "the agent can't read your data" — it's "the agent posted to your account, emailed your customer list, or refunded an order twice" at 2am with nobody watching.
X made the safe choice the default. When we wire an AI agent into your Shopify store, your CRM, or your phone system, we make the same call on purpose: reads are cheap and reversible; writes get a boundary and a human checkpoint. An agent that can look up an order is useful. An agent that can cancel one without asking is a liability.
The tooling is finally standardizing enough that this is a design decision, not a science project — which is exactly when it's worth doing right.
Key takeaways
- X shipped a hosted MCP server (June 30) — read/search only; the Write API is deliberately excluded, so agents can't post
- It connects Claude, Cursor, Grok, and other MCP tools using your own account permissions
- The MCP spec's July 28 revision makes the protocol stateless, load-balancer-friendly, and OAuth-aligned
- The operator lesson: give agents read access freely, but put every write behind a boundary and a human check
Thinking about connecting an AI agent to a system you actually run on? We build integrations where reads are easy and writes are guarded. See how we build or tell us what you'd wire up.
Sources: TechCrunch, Model Context Protocol blog.
- #mcp
- #ai-agents
- #automation
- #integrations
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.