Web MIDI / Bluetooth MIDI are both established apis… reliability is likely to be more dependent on your hardware and drivers than the ‘app’ that uses them. similarly, permissions are down to that api or OS, and AI support should be fine.
(ofc, AI can have hallucinations for any api, its never 100%)
Im not really sure what you are trying to do. that needs you to write/generate code…
“I want a simple companion app that reads controller data and shows mapped parameters”
sounds like something you could just get an app for, rather than create your own.
sure vibe coding is the buzz for now, but still many good reasons use something that someone else has not only created, but tested and supports.
if you want to go down this rabbit hole… thats cool too, it can be fun…
if you are using a ‘app builder’, and it generates code, then you could always use that as starting point, then point an LLM at the result to tweak it if necessary (for something the app builder doesn’t support)
however, I’d test this beforehand, its possible for code to be ‘obfuscated’ (made hard to read), whilst this doesn’t make it impossible for AI to process, it can make it quite a bit harder (= spend more tokens!) .
Im not sure this is common practice (for AI apps, it is for pro apps) at the moment, but I expect it will, given the popularity of vibe coding 
or as @NothanUmber said, just vibe code it from scratch if you have a clear idea what you want, scope is limited etc… just becareful, its easy to end up deeper you expect at the outset… which can be fun - if its something you enjoy (and dont mind spending tokens / $)
as for LLM,
I think local stuff is either very expensive, or pretty limiting.
Claude’s ‘cheap’ sub (20$) , I found too limiting to be useful… and from what I hear, even the $100 is easy to burn thru.
I personally (after copilots repricing) settled on opencode mostly zen, but also supports with byok… you could use opencode/go for simple stuff…
I use a variety of models, though Deepseek variants I guess is main one.
What I like about this is. you don’t hit session limits as you are paying by the token.
also, $20 is a lot of tokens if you are selective with models,
then factor in you are often paying for subs when you are not that active.
I think what I like most though, is model selection and ‘token’ management are good skills to learn. both for frontier models, and also later for local llms.
as @NothanUmber alluded too, I think the days of cheap LLMs and subs is limited,
Github copilot used to be crazy good value at $10/month, and now is much more expensive.
Claude, seems to have kept its price stable, but you get much less for it, as I said $20 i found almost unusable.
Not too surprising really, costs are going up, and also Id assume the corp market is going to be their main cusstomers - and they have AI budgets.
so unfortunately, I think hobbyists/vibe coder have started to get priced out the market.
possibly, local LLMs will be the future for them, but with hardware costs (RAM/GPU) at an all time high, probably not for next year or so…
but hey, ride the wave - its fun times 