OK, so there's a few things I've built. First, I want to clarify that there's the concept of a "tool call" in LLMs. i.e. the LLM generates a specific sort of response which is intended to be interpreted as a request to run a function (either in llama-server [if you enable that] or in your client). The LLM needs to be provided with a list of the tools that are supported (including the name of the tool, the description, and arguments) for it to know that it can call them.

Basically, you send a prompt, the LLM "thinks" for a bit, it spits out a tool call as the response, the client code sees the tool call, runs the appropriate function, then submits the result back as another prompt to the LLM and it continues on generating a response.

I've built custom tools (in my own custom client) for accessing limited parts of my file system in a particular read only fashion. (I do not trust these things to write to my disk autonomously...)

For example, I have a find_files tool that LLMs can call with a project name (which I need to provide -- restricted to a few specific locations based on a config file) and which basically runs a listdir and returns the results as JSON to the LLM. I also have a read_source_code tool that allows specifying a project and a file and it will return the text of that file.

This is technically redundant with features that are available out-of-the-box in llama-server but I'd rather run llama-server locked down in a container and provide it access with no more than I want it to have from code I control.

The custom client I wrote includes my own web UI (which is simpler than the one that comes out of the box with llama-server but is tailored to my specific tastes), and the code to run those custom tool calls. The first pass was also my first serious experiment with vibe coding; I wrote a very simple skeleton by hand and then iterated on it allowing the LLM to write most of the code so I could see what that process was like and how it would go wrong. As you might expect, that turned into a complete mess and I scrapped and rewrote pretty much the whole whole thing after a few weeks, but it was quite educational!