Homelabbers and Desktop Tweakers may want to move beyond just plain AI chat
The days of asking AI a question, or pasting a config file in, and then pasting (or typing) the updated file back, are no longer needed. I have over 47 running docker containers, with almost as many stopped, and when I want to create a new docker stack I'm never sure if there will be a port clash. Now I can just ask an agentic AI to check all my stopped and running containers and fix any port clashes, restart containers, diagnose why something is not working, check reverse proxy arrangements, etc.
Agentic AI is a term often heard, but in simple English, it is AI that actually does stuff, not just answer questions. Sometimes there is an MCP (Model Context Protocol) server to install, or a connector, or a plugin. Those mechanisms give selected access to the AI to actually do tasks like reading and updating a config file itself, check log files and advise on corrective actions (or even do those actions), etc.
A chatbot AI answers questions, but an Agentic AI is a combination of AI with carefully selected tools installed, and permissions that must be granted by the user.
In my earlier example with Docker Portainer, I put some code into Claude's config to connect to my Portainer instance on the LAN directly. In this way it can read all my existing container configs and learn what style I'm already using, and create something that really fits my environment with its unique volume mappings, etc. I also gave it directory access to my Conky config files (and the scripts folder for that) so that it can tweak any changes I require.
I have also installed the n8n connector which has taken my breath away. The n8n service is workflow automation tool (for example I use a n8n workflow for checking my blog RSS feed, and then automatically pushing my posts out to the various social media accounts I post at). That is usually a lot of back and forth as I have to fix small issues. Now Claude uses the n8n connector and literally creates a n8n workflow from scratch for me. Yesterday it created a workflow to search my Nextcloud Notes content, retrieve it, update it, add new notes. It tested and figured out what did not work, and applied the fixes all by itself (with me just reviewing and clicking apply for the various write actions).
So many possibilities open up, but there are some safeguards worth keeping in mind:
* Most such connectors have permissions that default to ask first. You can go a step further by granting always read permission, and keeping ask before writing, or if you are really living on the edge then permission granted for read/write/delete/overwrite automatically (and only for trusted tools).
* I have told Claude to remember to always back up any config file that it is going to write to or edit, but more importantly I have my own automated and independent backups that I can restore myself if I have to.
* Filesystem connectors require you to add specific paths first. Do not add the root folder, but stick to just specified folders you want help on.
* The Portainer access for example cannot access the .env files where typically passwords and API keys are stored (in Claude's case for this connector so you may want to check on what you use for this).
* The cloud side cannot reach inside your network or desktop without a locally-running connector. So when you want the AI to reach your LAN, filesystem, Docker socket, localhost services, etc you need to install an appropriate local connector/bridge service.
* Ask to review differences before applying changes.
* Don't give an AI access to secrets merely because a tool happens to be capable of reading them. You can often put secrets into a file elsewhere, where your scripts can put them in themselves, but AI has no access to read that folder.
But start small and just give it access to one useful thing. This lets you see what it can do, and then you can decide whether you want to give it more. Even a few small automated tasks end up saving a lot of time down the road.
And yes, some AI tools are better than others at some tasks. So when I asked Gemini which would be best for my types of tweaking and configs that I do in my home lab, it recommended Claude without hesitation. Chat is still also very useful for asking questions about setting up your Agentic AI capabilities, and what safeguards you should apply.
I decided to also start paying for AI now as it is performing a truly useful function, and saving me money not only in time now, but also to better utilise the free self-hosting services that I have. I have already saved more than $20 per month on other subscription services (see my video about what Wallos did for me —
https://www.youtube.com/watch?v=_9F4szRS6mY).
It is not to say AI is perfect (it still needs human supervision) but it is undoubtedly a super tool to help make humans more productive themselves. For the same reason humans needed to master the use of the wheel, levers, early cars, telephones, and Internet search engines, if we don't do the same with AI, we could lose our jobs to other humans who do use AI effectively as a tool. Cars and telephones never directly threatened jobs at all, in fact they created many new types of jobs, but those who harnessed those technologies gained leverage over those who just ignored them. Whether we like it or not, the world has moved on to another level. The great advantage we have with this new tool is, we can even ask it how we can use it better and more safely (just check its answer elsewhere too!).
The more worrying side we will need to solve, is how AI is being weaponised for hacking attacks, and even more dangerous, for creation of convincing AI-generated misinformation that can destabilise whole countries. This is something we need to solve, because cars, phones, the Internet, etc are also all used for bad purposes too, but they don't just get cancelled.
#technology #homelab #selfhosting #AI
