upvote
MCP is a tool more for security than anything else. If you give your agents access to an API key, there's a chance they can accidentally or maliciously leak that key. If the MCP server has access to the keys instead, it takes that possibility away. That's not always something you need to care about, but it is very important for some people's threat model.
reply
> MCP is a tool more for security than anything else.

The agent can get to the resource through the MCP server or using API key. I personally do not see the benefit MCP is providing here. Sure you can reduce the exposed surface at MCP layer, but I do that at the API layer. I don't need to add another layer here.

I can kinda understand if you do not have control of the API layer and/or you have to expose the API layer to the public Internet as well. Most of the time that is not the case for me.

reply
The idea is that with an MCP, the agent doesn’t have access to any credentials. It can’t leak an API key.
reply
There are many ways to do this without using mcp, for example one of the many proxy servers that inject the token into requests. I use a custom solution that redacts all secrets to agent transcripts before the agent sees them in a hook.

I find it way better to be able to confidently tell agents to use CLIs than worrying about partially implemented mcps that need configuration and are often yolo’d with npx @latest anyway

reply
Sooo, how does the agent authenticate with the MCP server?
reply
My own approach is to put the API key in the MCP server, apply principle of least privilege there, and firewall access to MCP with agents being on same private network with Tailscale.

Not only giving an API key to an agent can leak to the model because of harness issues or too broad reading rights, but also most providers don't give the ability to apply principle of least privilege to an API key.

I don't want to give an agent full R/W access to any of my services/accounts.

reply
The agent doesn't, the harness does. It's separate from a normal conversation / agent runtime environment. How do you suspect auth keys can leak from an MCP that's been added to chatgpt.com?
reply
With a different key, obviously. This lets you keep the MCP server firewalled to your local network only while still allowing home assistant itself to access the internet
reply
You give it a key of course! And if that doesn’t fit your security model and threat vectors, simply give the key to another MCP.
reply
Verging away from the topic, but I've found Claude is a great addition to HA. I want smart home features, but don't have the time/inclination to learn HA's way of working. Historically I just defaulted to Google Home because it was easier, but with Claude I barely need to touch HA configuration at all.

Recently I wanted to set up a slightly complex routine involving some lights and a couple of motion sensors. It feels like magic to be able to describe the behavior I want, briefly discuss the implementation, and walk past the sensor and see it in action.

reply
One of the key points of the comment you replied to is that you don’t need a very powerful model to do MCP stuff, such as local Qwen. The “let the agent figure it out” approach works much better with powerful models, such as Claude (which you mentioned you are using).
reply
So MCP is useful until cheap models get better?
reply
I +1 this approach. We are doing similarly. Instead of providing MCP, we are simply providing API key and documentation. Folks are able to then just paste that link and use our service.

User will interact or build apps with simply text like: "Give me all the issues that are X, context: https://somedomain.com/llm.txt"

llm.txt will have all the API instructions

reply
That's smart! That reminds me, I have to get back into HomeAssistant. I had my whole house through it a few years ago, but the complexity and things breaking in hard to debug ways made the experience too frustrating for my wife and visiting relatives.

Having an agent keep an eye on stuff and fix things proactively should make the experience much better. Plus I can no longer write yamls at last.

reply
Developing for HA with Claude has been great. Not only can it make all of those yaml changes based on natural language goals, but it's so easy now to create a custom dashboard or configure various apps. I was struggling with both the ChoreOps docs and its fairly cumbersome interface until I pointed Claude at it and told it what I was trying to do.
reply
> No MCP needed

Definitely helps that the home assistant api is documented online most likely in the training data.

reply
One of the nice things about MCP is that the tools have descriptions that provide guidance to the LLM. You can provide documentation in other formats like OpenAPI but its nice that the documentation is so closely coupled with MCP servers.
reply
> One of the nice things about MCP is that the tools have descriptions that provide guidance to the LLM.

It is more of a curse than a blessing. MCP pollutes agent context even when you are not using it. Use a manually invoked skill instead if you don't want to be wasting tokens on every turn and bloating up agent context making it dumber in the process.

reply
> MCP pollutes agent context even when you are not using it

MCP in some harnesses bloats context. MCP in some harnesses doesn't bloat context.

reply
How are you interacting with Claude for this?
reply
I just run claude's cli tool?
reply
Same question… but I would assume he’s running Claude Code on a PC that is on the same network.
reply