ChatGPT Sites can now host MCP servers. On October 1, 2026, two OpenAI product leads publicly announced on social media, one after the other: users can build and deploy MCP servers directly inside ChatGPT, turn them into plugins for use in ChatGPT, and keep access private, share it with specific people, or make it public. For MCP developers long tormented by "finding a host, setting up certificates, staying online," this is effectively the official, ChatGPT-native hosting answer that removes the need to run your own server.
Getting the facts straight
This was not an official blog post or press release, but a direct announcement by the product leads on social media. Tibo Sottiaux, head of Codex, posted that "you can now build and deploy MCP servers right through ChatGPT," emphasizing that access can be restricted or shared publicly. Max Stoiber, head of Plugins, followed up to confirm that ChatGPT Sites can host MCP servers, including plugin extensions.
The example flow is concrete: ask Sites to "create a to-do list I can use in ChatGPT," and behind the scenes it runs the full chain of "create the MCP server → deploy to Sites → convert into a plugin." A third-party hands-on guide published the same day filled in the key details: save a version before deploying, every deploy URL is the production environment; sharing goes through Sites audience controls (only you, invited people, workspace, or public); but listing in the public plugin directory is a separate review path — sharing on Sites does not mean entering the public directory. Readers new to MCP can start with this explainer on how MCP differs from APIs and plugin calls.
Why the "host" was the first thing to get solved
The MCP protocol itself isn't complicated. The real barrier was always the "host": a server that's continuously online, speaks Streamable HTTP, and is reachable by ChatGPT. For indie developers and small teams, writing the tool logic might take an afternoon — finding a VPS, configuring HTTPS, and guaranteeing stable uptime is the step that drives people away. That's exactly where OpenAI struck this time, pulling the most annoying operations work into ChatGPT's own territory.
Who gets it first, and where the boundaries are
Two groups benefit most. First, users who "can describe a need but can't operate a server": until now they had to wait for others to publish plugins; now they can build their own dedicated little tool right in the conversation. Second, teams building internal tools: Sites audience controls are a natural fit for "team-only" scenarios, with no need to maintain a separate auth and release pipeline for a small tool.
But the boundaries are equally clear. Sites is currently in public beta, covering the Plus, Pro, Business, Enterprise, and Edu plans, each with usage limits; scenarios requiring direct connections to private data sources, raw TCP, or sensitive data and data-residency requirements still need self-hosting. More importantly, the public Sites documentation is still written as "website hosting" — official MCP hosting documentation hasn't caught up yet. For production-grade or sensitive services, testing demos and internal tools first is the prudent move; going straight to production is premature. To compare with the other camp's approach, see Claude opens an official plugin submission channel: MCP connectors and Agent Skills can go straight into the directory: Anthropic is betting on "open directory review," OpenAI this time on "native hosting" — the two routes complement each other.
The real signal
MCP is shifting from a "developer protocol" to a "capability layer users can directly consume." When the barrier to building an MCP server drops from "can operate" to "can describe," the supply side of tools opens up completely — what matters next is no longer who can set up a server, but who describes needs most precisely. Hosting itself has become a new contested prize: third-party MCP clouds, the Claude connectors directory, and now OpenAI's Sites hosting are all chasing the same thing: becoming the "default landing spot" for agent tools. OpenAI's strength is its closeness to users; its weakness is documentation and ecosystem maturity that haven't caught up yet. This game has only just begun.
