1. Bottom line: Cloudflare is at the door, but ChatGPT’s door handle is not visible
The setup looked promising: install the Cloudflare Plugin, turn on Developer mode, and surely ChatGPT should now be able to inspect DNS or Analytics. Instead, the Cloudflare page showed Skills such as Agents SDK, Workers, and Web Performance. Trying to invoke @Cloudflare in chat produced no usable app. The package arrived, but the box contained documentation rather than the cable.
This does not mean Cloudflare lacks MCP support. Cloudflare runs official remote MCP servers with OAuth access to Cloudflare accounts. The likely bottleneck is exposure on the ChatGPT side: the plan, role, workspace, model, or current UI may not expose the entry point needed to create or invoke the actual app connection.
2. What happened: “Plugin installed” does not mean “external account connected”
In 2026, “Plugin” is a broader container than many users expect. A plugin can package Skills, Apps, and app templates. An App is the part that actually connects ChatGPT to external data or actions.
In this case, the Cloudflare listing visibly contained Skills, but no underlying Connect/App surface appeared. That is how you can have “Cloudflare Plugin enabled” while ChatGPT still cannot read the account’s DNS or analytics. It is the software equivalent of bringing home a brochure that says “Wi‑Fi ready” and wondering why your router has not paired itself. The brochure is fine. The router is fine. The pairing screen never appeared.
3. Plugin, Skill, App, and MCP in one pass
A useful mental model is: Plugin is the storefront/package, Skill is the manual or domain guidance, App is the real integration port, and MCP is the common protocol connecting AI clients to external tools.
- Plugin: a package used to discover and enable workflow capabilities in ChatGPT/Codex.
- Skill: guidance, best practices, or specialized knowledge for using a service.
- App: an integration that connects ChatGPT to external data or actions.
- MCP: a standard protocol for connecting AI clients and external tools—the AI equivalent of a common connector.
The confusion comes from all four appearing under the same vendor identity. “Cloudflare knowledge Skill” and “connect to my Cloudflare account through an App/MCP” are not the same thing.
4. The July 9, 2026 Plugin migration made the naming even stranger
OpenAI says that on July 9, 2026, discovery moved from the App Directory to the Plugin Directory. Plugins became the primary discovery surface for workflows across ChatGPT and Codex. Apps did not disappear; they remain the integrations that connect to external data and actions, while plugins can bundle apps, skills, and templates.
OpenAI also explicitly says publishing a custom MCP app does not automatically publish a separate plugin. The reverse implication is equally important in practice: installing a plugin does not guarantee it contains an account-connected app. Same Cloudflare logo, same vendor name, “official MCP” in the description—then you open it and find Skills. It looks like an Ethernet jack and turns out to be the instruction manual for Ethernet.
5. Cloudflare’s side is ready: its official MCP reaches more than 2,500 API endpoints
Cloudflare operates managed remote MCP servers. Its Cloudflare API MCP server can access more than 2,500 endpoints covering DNS, Workers, R2, Zero Trust, and the wider Cloudflare API through two tools, search() and execute().
On July 28, 2026, Cloudflare updated its MCP servers for the MCP 2026-07-28 specification and told new clients to use the /mcp Streamable HTTP endpoint. OAuth is a supported connection model. Cloudflare is essentially standing at the entrance saying, “Yes, AI clients may enter.” The remaining question is whether ChatGPT’s elevator currently stops on this floor for your account.
6. OpenAI’s current documentation looks slightly split-brained
This is the most interesting and most important caveat. OpenAI’s developer-mode documentation describes Apps, full MCP, and developer mode primarily in the context of Business and Enterprise/Edu. Yet the FAQ on the same documentation says Pro users can connect MCPs with read/fetch permissions in developer mode, while full MCP write/modify remains limited to Business and Enterprise/Edu.
So the text says Pro read/fetch connectivity exists, but some real UIs still do not expose Apps or Create. OpenAI separately notes that app availability depends on plan, region, workspace, role, model, and interface. Therefore it is unsafe to conclude either “you configured it wrong” or “it must already work for every Pro account.” The docs and the UI are on the same train, but apparently in different cars.
7. “It feels close” is a reasonable inference—but there is no arrival timetable
There are real reasons for optimism. OpenAI’s FAQ explicitly mentions Pro read/fetch MCP. The Plugin Directory migration happened in July 2026. Cloudflare refreshed its MCP implementation at the end of July. Developer mode is already visible in some individual environments. Many of the pieces exist.
That makes this look less like a distant concept and more like a distribution/UI exposure gap. But none of that proves it will appear next week or next month. Beta rollouts do not ship with railway timetables. The platform has installed the departure board; the train time still reads --:--. Please hurry.
8. What changes once it works: website operations become “ask the system” instead of “open fourteen dashboards”
Given the scope of Cloudflare’s API MCP, a connected setup could answer operational questions such as: What are the DNS records? What is configured in Workers? Did errors increase? What does this zone look like? Analytics and observability MCP services can further bring logs and traffic data into an AI workflow.
The larger payoff is operational chaining. Check GitHub changes, deployment results, Cloudflare errors, and the public HTTP state in one flow. Instead of a human opening a warehouse, factory, logistics center, and storefront one tab at a time, the human asks, “Anything broken today?” and the AI walks the chain. Website management moves from tab ceremony to talking to the factory manager.
9. Read and write are different worlds: do not treat OAuth and permissions casually
MCP convenience does not eliminate access control. OpenAI currently frames Pro custom MCP around read/fetch, with full write/modify MCP reserved for Business and Enterprise/Edu. Cloudflare’s MCP can support changes, but the actual ability to modify data depends on the ChatGPT plan, app configuration, permissions, and confirmation flow.
Do not paste API tokens directly into chat text. Prefer OAuth with narrowly scoped permissions, or store minimal CI/CD tokens in secret storage. Before telling an AI to “handle everything,” decide what it is allowed to read and what it is allowed to change. You would not give a new employee permission to delete every DNS record on day one. The employee being silicon does not improve that idea.
10. What to do now: stop searching for a button that is not rendered, and be ready when it appears
If Apps/Create is not visible, spending three hours searching for a way to click the nonexistent button will not cause it to spawn. Cloudflare’s dashboard cannot push a ChatGPT app into an interface that does not expose the capability. Keep Developer mode enabled and watch for the relevant app/custom MCP surface.
If the operational need cannot wait, GitHub/Codex, CI, or another MCP-capable client can connect to Cloudflare API/MCP first. When ChatGPT’s UI path becomes available, the same operational model can be moved into conversation. The factory can already run; what is missing is the direct phone line from the executive office to the factory manager.
The conclusion is simple: Cloudflare already supports MCP. ChatGPT is actively developing MCP. A Cloudflare Plugin may already be visible. But “a Plugin exists” is not the same as “my Cloudflare account is connected through an App that this UI can invoke.” The manual arrived. The connector has a standard. The hole in the wall is still missing. Please hurry.
