ChatGPT MCP + Home Assistant via Cloudflare Tunnel – Connected but no command access

[KinoFuN]

Hello,

I apologize for my English; I am using a translator.

We are trying to use ChatGPT with Home Assistant via the MCP (Multi-Client Proxy) server. We have followed the setup instructions step by step.

What we did:
1. Tried on both MacMini and Synology Docker with the latest versions.
2. Installed uv and cloudflared as described in the documentation.
3. Created a Cloudflare Tunnel to our Home Assistant instance, which connects successfully and shows the public URL (https://travels-imperial-wrote-tough.trycloudflare.com/private_4ef536845333d9f8cb609616).
4. ChatGPT successfully recognizes the connection to Home Assistant. All services appear as public.

Problem:
• When sending any command from ChatGPT, it responds that it does not have access.
• In the Home Assistant connection info, the following fields appear:
• Supported authorization: None
• Used authorization: None

It seems that even though all services are public, without a valid token or proper authorization ChatGPT cannot execute commands.

Question:
• Are we missing an authorization step or token configuration for ChatGPT to be able to send commands to Home Assistant via MCP?
• Is this expected behavior for public services, or is there a setup step we overlooked?

Any guidance or tips to make ChatGPT able to control Home Assistant via MCP would be very appreciated.

[julienld]

In ChatGPT.com, you have to be in "Developer mode" for this MCP to work (orange border around the chat window). Are you in that "Developer mode"?

PS: You should not post your private url in a public discussion.

[KinoFuN]

Hi,
first of all, thank you very much for your reply and for taking the time to look into my issue. I really appreciate it. To be honest, I don’t have anyone around me with whom I can discuss these kinds of technical topics, so any guidance from the community means a lot to me.

Regarding Developer mode:
Yes, I am using Developer mode.
I was in Developer mode both when creating the ChatGPT ↔ Home Assistant connection and when testing command execution directly from the ChatGPT interface. I am aware of the visual indication (different border/color) and can confirm that I was in Developer mode during these tests.

Regarding the public URL you mentioned:
I understand the warning and agree — the URL I shared is temporary and test-only, and it will of course be changed.
I intentionally shared the full context of that window not because of the URL itself, but to highlight something else that I found important:
• Supported authorization: None
• Used authorization: None

This was the main reason for my suspicion that the issue might be related to authorization or how MCP / ChatGPT accesses Home Assistant, rather than the security of the URL itself.

Additionally, I would like to ask about one more thing:
During my testing, I also tried using a permanent Cloudflare Tunnel, not only a temporary one. The behavior was the same — the connection appeared to be established successfully, services were visible as public, but when trying to execute any command, ChatGPT reported that it does not have access.

So my question is whether, in your opinion, part of the problem could still be related to the SSE / Cloudflare Tunnel / authorization mechanism, even if the MCP server itself behaves correctly.

Thanks again for your response, and I appreciate any further guidance.

[Shadow941]

same problem. Connected, but no command access

[KinoFuN]

Sorry in advance if the wording sounds a bit odd – English is not my native language and I’m using a translator, but I wanted to share my experience in case it helps.

I had exactly the same situation: connected, HA actions visible, but “no command access”. In the end this turned out not to be an authorization or MCP bug, but a combination of expectations and a few setup details that are not obvious at first.

First, “Supported authorization: None / Used authorization: None” is NOT an error for ha-mcp. This confused me the most. ChatGPT does not authenticate to Home Assistant directly. ha-mcp authenticates to HA internally using HOMEASSISTANT_TOKEN, and ChatGPT only sends tool invocation requests to MCP. This does not mean missing permissions, missing OAuth, or missing API keys. It is simply how MCP is designed. If authorization were actually broken, you would see 401/403 errors or no HA actions at all.

Developer mode is mandatory, but not sufficient on its own. The maintainer is correct: without ChatGPT Developer mode (orange border), MCP tools are never called, only described. However, even in Developer mode, ChatGPT will not automatically call Home Assistant unless the command is explicit.

I also lost quite a bit of time around Cloudflare Tunnel and certificates. In my case the tunnel itself was working, but I initially ran into problems because I tried routing MCP through Synology and hit TLS/certificate issues, and I also had an incorrect DNS setup in Cloudflare. What finally worked reliably (all on the free tier) was a permanent Cloudflare Tunnel, a correct DNS CNAME pointing to the tunnel ID with proxy enabled, no extra reverse proxy in front, and clean TLS end-to-end between Cloudflare and MCP. Once this was fixed, the MCP connection was stable.

The key insight for me was that ChatGPT does NOT call Home Assistant automatically. For every message it decides whether it can answer from its own knowledge or whether a tool call is clearly required. For example, “What’s the temperature in the kitchen?” may be answered without calling HA, while “Use Home Assistant and read the temperature sensor in the kitchen” will invoke MCP. The same applies to actions. Vague requests may result in a text response, while explicit action-style commands trigger real calls.

When everything works correctly, you will see a confirmation dialog (ToggleLight / Cancel). After confirming, a real service.call is executed in Home Assistant. I’m attaching a screenshot of that confirmation dialog from my setup.

For me, the final proof was using an explicit command like:
Turn off the light “Led pás kancelária”.
ChatGPT showed the HA confirmation dialog, I confirmed it, and the light physically turned off. At that moment it was clear that MCP, authorization, and the tunnel were all correct; the missing piece was command clarity, not permissions.

TL;DR: “Authorization: None” is OK, MCP does not auto-call HA, Developer mode is required, and you must use explicit, action-style commands. If you see HA actions and a confirmation dialog, your setup is working. These were exactly the details I was missing at the beginning, so I hope this saves someone else the same confusion.