I was just trying to start to play with MCP and homey but somehow i can’t access mcp.athom.com anyone else having issues opening the URL?
Claude adds more to the free tier, but does not seem like Custom MCP is there still?
I always get compaction error from Claude when trying to create advanced flows
I connected a free LLM when homey MPC was released. I dont remember which. Where in my homey account can I see what services are connected to my homey and delete them?
That’s not possible, you can log out of all linked services by changing your Homey account password.
I would think when adding a new service it would create a kind of app password/registration and not be linked to your personal password - that would not be much professional by a company like homey
I tried to connect Claude with Homey today, but unsuccessfully.
everything seems going smoothly, but I can never get the “connected” status in the Claude connector page.
interested to know if someone had the same issue and found a solution.
I have a self hosted instance of Homey. I connected the MCP to my app using my personal access token and was able to connect and retrieve endpoints. However, when I perform any action, I always get an empty list of homeys. However, it’s telling me my account’s email address, so clearly it’s connected. Any idea? Is this a limitation of self hosting?
I wanted to share a project I’ve been working on: connecting my Homey Pro to Claude AI (Anthropic) using the Model Context Protocol (MCP).
The goal was to move beyond simple voice commands and create a truly intelligent collaborator that understands my home’s technical profile, energy goals, and device relationships in real-time.
The Technical Architecture: Purely Local
One of the key priorities was privacy and speed. Instead of relying on cloud-to-cloud integrations, this setup uses a Local API approach:
-
The Bridge: A custom MCP server running locally on Node.js.
-
Direct Connection: Claude communicates directly with the Homey Pro via its local secure endpoint (
https://[IP-with-dashes].homey.homeylocal.com/api/manager). -
Authentication: High-performance local session tokens are used to authorize requests, ensuring no data leaves the house unnecessarily.
The “Home Brain”: Markdown Context Integration
What makes this setup unique is that I’ve fed Claude a comprehensive Home Profile in Markdown format. This serves as the “long-term memory” for the AI, containing:
-
Energy Certificate and energy consumption Data: All technical values from my official energy certificate (e.g., heat loss coefficients, predicted annual consumption).
-
Technical Specs: Details about my Air-to-Water Heat Pump, ventilation system, and structural insulation.
-
Maintenance History: Operating manuals and service logs.
Real-World Capabilities:
By combining real-time device data from Homey with the static technical data from the profile, Claude can perform advanced tasks:
-
Efficiency Audit: Claude can pull real-time power consumption from my heat pump and compare it against the climate-adjusted values in my Energy Certificate to see if the system is operating at peak efficiency.
-
Smart Maintenance: Instead of simple timers, Claude can suggest actions.
-
Proactive Optimization: Claude analyzes sensor data across the house to find discrepancies between my home’s insulation profile and actual heat retention, suggesting adjustments to my heating curve.
Why this matters:
Having an LLM that doesn’t just “see” your devices but actually “understands” the physics and technical requirements of your specific building is a game-changer. It turns a smart home into a truly managed home.
I’m no coder or app builder, so I used Gemini AI to build this connection… a fun project I might add.
Sounds interesting.
Thanks for sharing the concept of your project.
I’ve been experimenting with the Homey MCP server together with ChatGPT and overall it works surprisingly well for basic interaction with my Homey.
However I noticed a few limitations and I’m curious if these are current limitations or things that might be expanded in the future.
Advanced Flows
It seems that MCP can only see Advanced Flows that contain a Start block. Flows without a start trigger do not appear. Is this expected behaviour?
Querying historical data
Another thing I was hoping to do is ask questions about my home, for example:
-
“What was the temperature in the living room yesterday evening?”
-
“How often did the heating run today?
At the moment it seems I can only access current device states, but not historical data from Insights.
I’m using MCP together with ChatGPT and it already opens up a lot of interesting possibilities. Being able to query historical data or ask questions about the house would make it even more powerful.
Short hint: you can also use Homey MCP with new Windows and Mac Codex app, connecting it with your ChatGPT account.
https://openai.com/de-DE/codex/
Advantages here: Higher context window. The new GPT-5.4 is very great at managing tools from mcp.
AI model benchmarks: https://artificialanalysis.ai/
Alternative way for all models: https://opencode.ai (also have a desktop GUI app if preferred)
I’ve been using Claude (pro) to rewrite a bunch of convoluted advanced flows together with me, with varying results.
I really need to check and double check all suggestions it makes. Often I think I spot a problem, mention it, and get an answer like ‘Yes, indeed, I overlooked that’ or something similar. Or I suggest a simpler solution to a problem, it responds with something like ‘You are right, that is a more elegant solution’.
Last weekend I spend hours with Claude rewriting things to circumvent a certain problem. At the end, the problem persisted. Claude said ‘It seems it doesn’t work because of ’. I reminded it we were rewriting all of this to solve . It seems it can lose its goal sometimes in the progress.
I’m just an amateur with al of this, so I’m sure there are ways to tweak Claude and improve its quality. Right now I use it inside a project called Homey, and stated some basic rules. I just keep feeling it lacks ‘understanding’ what it is doing, and how advanced flows work.
I’m curious to learn how other people instruct Claude (or any other ai) to behave.
Hi,
I’m reposting this because I deleted my previous topic after I accidentally included a Homey identifier in it. 
Context: I’m currently testing read-only connector stability before using AI to generate a flow for my ventilation system. I am not creating or modifying any flows yet. At this stage I only want to verify that the connector can complete a sequence of read calls reliably in a single session.
What happened:
-
list_homeyssucceeded -
list_flow_trigger_cardsfailed immediately
Exact error:
flow.trigger.autocomplete.getArgumentValues is missing in manifest
I stopped after that first failure.
Important context:
-
read-only test only
-
no flow creation
-
no flow update
-
no flow deletion
-
no assumptions beyond the live call results
My questions:
-
Is this a known issue?
-
Does this usually point to a broken app manifest or Flow card definition in one installed app?
-
Can one faulty app break
list_flow_trigger_cardsfor the whole Homey session? -
Has anyone seen this specifically with autocomplete arguments on trigger cards?
I’m mainly trying to understand whether this is a connector stability issue, a Web API issue, or an installed app / manifest issue before I move on to the next stage.
Thanks.
Can you show the actual code? The reason I asked in your previous question if it was AI-generated is because I got the impression that your AI hallucinated (part of) the code.
Sure — but to clarify, there is no custom Homey app code involved on my side.
What the AI said it did 
was a read-only connector test against an existing Homey connection, before attempting AI-generated flow creation for my ventilation system. The test was simply a sequence of live read calls in one session.
The sequence was:
-
list_homeys -
list_flow_trigger_cards -
list_flow_condition_cards -
list_flow_action_cards -
list_devices -
list_flows
What actually happened was:
-
list_homeyssucceeded -
list_flow_trigger_cardsfailed with:
flow.trigger.autocomplete.getArgumentValues is missing in manifest
So this was not generated Homey code that the AI wrote and ran locally. It was a live connector/API read test, and I stopped immediately on the first failure.
That said, your suspicion is fair. My current question is exactly whether this points to a real Homey/app/manifest issue or whether the connector/AI layer may have interpreted something incorrectly.
I can also post the exact raw request/response shape from the failing read call if that helps narrow it down.
Ah okay. Are you using an MCP server on Homey?
Yes… any advise is welcome 
Let’s start with which MCP server?
