I want to start getting into LLMs in HA and using MCP servers with them. I setup the OpenAI integration and can use it in my (voice) pipelines. Now I’d like to enhance the LLM with additional features like checking my calendar. However, I am not really sure where to start and I didn’t find the tuturial/post to answer my question.
I’d appreciate if someone could point me out in the right directions regarding these questions:
I see that there are a lot of mcp servers out there. However I am not sure how to run them. Take this one as a simple example server.
Where do I add the mcpServers config? As I understand it this would be added to something like Claude running in a desktop application, right? Can I use this in the HA server environment? Can I use such a server when using a cloud model (as I don’t have a GPU available at the moment)?
When I have figured out how to run that server I have to use the mcp integration to add is as a tool and add some instructions to the prompt, right?
I found n8n to be a pretty interesting tool that allows AI agents to interact with other applications. So there are multiple ways to add mcp-servers (through running them as described above and using n8n), right?
I found this post in the community forum regarding this and am wondering whether this is the way to go.
TLDR: Ideally I’d have an LLM available in HA that is able to call other applications, how do I do that?
First you need an mcp server that supports remote connection. There is no standard way to do that but usually you’re running some kind of container that has your ‘server’ in it that server / service will have a way to connect and needs to support SSE for remote connections so the HA MCP client can even connect. From there it’s usually OAuth.
Cloud flare has a blog article of I can locate it.
Now as you can see I’m painting a very self-host or BYO service picture… Are you really looking specifically to use MCP (it’s hard right now and may not be what you are actually looking for)
Or are you trying to add capability to your llm because - if its that you should learn how HA uses tools before throwing MCP on the mix.
Don’t get me wrong MCP is awesome. Just it’s not end-all. There are different methods and your end game dictates the path. The limits of the system dictate what paths are available.
You have a lot of prompt and context to get through before dealing with MCP…
See Friday’s Party for details if youre interested…
Or are you trying to add capability to your llm because - if its that you should learn how HA uses tools before throwing MCP on the mix.
Yes, this is what I am trying to do: Adding capabilities to the LLM of OpenAI. I just thought this would be done through mcp servers. I will read the linked post from you and see if I have any questions afterwards