Yes, yes, AI/LLM bad… anyway: much like the rest of the world, I have been trying to navigate this AI/LLM explosion. There is a lot of awful to it all and a lot of utility. I tire of the AI doomers and AI hype crowds (as well as corps pushing it into absolutely everything). There are so many posts about how AI bags up puppies and beats them to death or posts about how AI will make you a bajillion dollarydoos while you do a bunch of nothing. I think the truth lies somewhere in the middle.

Over the last six months or so I have been picking up various AI tools, mostly LLMs with some MCP integrations. It puts me in a state of cognitive dissonance. It’s often faster than I am a some programming tasks and seems to distill information pretty rapidly, too. I am a bit embarrassed when it writes better code than me and I don’t often talk to my colleagues about my AI usage (or theirs).

I do believe a world of bespoke and one-off custom software development is upon us. I, myself, have used LLMs to bang out scripts that would be annoying for me to write or to handle system-level tasks that bug me. I check its work and move on. I don’t know that I’d be able to do this very efficiently if I wasn’t already steeped in code/tech/systems/whatevers.

With the navel gazing out of the way, here’s the AI services I’m currently using and how I use them:

Kagi Assistant

I am a big fan of Kagi Search and I yammer about it pretty often. I am terrified that they’ll enshittify eventually but for now they’re great. Their Assistant tool provides me access to several models in my browser and I find it very useful.

I use the assistant when I have complex questions in order to batch search results. Often for technical queries where I don’t need the agent to have integration with the system but also personal queries. I could probably live without this but it’s definitely made finding information much easier. I really like that it provides references to everything and I use it a lot like a search engine that I can send queries structured as paragraphs to.

Claude

The elephant in the room. I have been making heavy use of Claude Code for system level tasks in recent history, especially at home. I added the linux-mcp-server to it and I have it handle sysadmin taks for me (with oversight). I don’t do this at work, only on my home systems. I am often worn out from dealing with other people’s infrastructure so having my own robot sysadmin is immensely useful. I also use it for sifting data and moving large batches of files around. Yes, I have good backups. I want to move this to a pile of OpenTofu and Ansible playbooks, though.

I also use this Claude account in the Zed editor with the context7 MCP for programming tasks. I mostly use the Opus model and I treat it like another engineer as if I were pair programming. I’m not often having it generate code but I have it work through solutions with me and help me sift docs or find function calls.

Zed Pro

My Zed Pro subscription provides me access to several of the major models and it’s mainly used as an overflow for me in the rare instance that I hit a limit with something else. If I did have Claude generate any plans or write any code, I might have GPT by way of Zed Pro take a pass at it after I do just in case I missed something. I do not make use of the Gemini models this provides because they have not produced good output for me. I use this service strictly for limited programming tasks as API usage is quite expensive.

LM Studio

I am experimenting with local LLMs but given my lean hardware they are slow and not very performant even with context7 and linux-mcp-server. I only use them for very precise tasks or if I’m just noodling around.

Lumo from Proton

I use this assistant for long form personal questions as it is allegedly private. Typically those questions are formulating workouts for me or helping with searching for general life information. I use it infrequently. I’m not sure that its results are good for technical work matters.

Summary

Overall, my AI/LLM usage is heavily technical and I’m not certain that it would work out if I haven’t already been dealing with technical matters for 30 years or more. These services benefit greatly from MCPs when I have the opportunity to do so. I also haven’t moved into any real general AI usage beyond searching and information gathering. I think in those areas, references and an audit trail are immensely important.

I’m on the hunt for ways to tailor my usage a little better, espeically with technical tasks and to have tighter integration with my tooling. However, the expense of the services, the available resources, and the rapid pace that all of this is evolving at, make it challenging for me to know where to start.