Somewhere around late 2024 I had 14 AI tools open simultaneously. A coding assistant, a model runner, two different chat UIs, a vector DB, and a few things I genuinely can't remember why I installed.
By mid-2025 I'd cut it to four.
I wrote about what survived and why over at Telegraph: https://telegra.ph/The-AI-Tooling-Stack-for-2026-Whats-Actually-WAlex-Morganorth-It-05-07
The short version: most AI tooling adds surface area, not capability. The tools I kept either (1) reduced the number of things I had to think about, or (2) were so low-friction they didn't cost me anything to keep running.
The ones I dropped failed on one of two criteria: they required active management, or they needed me to change how I worked to fit them instead of the reverse.
A few things I noticed after trimming:
My actual output went up. Not because I found a better tool — because I stopped context-switching between tabs trying to figure out which tool to ask.
Local inference is underrated. Running a smaller model locally for routine tasks costs you nothing and removes the latency tax on quick questions. I use it for code review drafts, summarizing meeting notes, and anything where I don't need frontier-level reasoning.
The "wrapper problem" is real. Most AI products are just thin wrappers around the same underlying APIs. If the wrapper doesn't add something meaningfully different — better memory, domain-specific tuning, an actually useful UI — it's not worth the extra dependency.
I'm now skeptical of any tool that requires a monthly subscription to justify its existence. The best tooling I use either runs locally, has a genuinely generous free tier, or is open source.
None of this is a recommendation to copy my stack. Tooling is personal. But if your current setup feels heavy, that's probably information worth acting on.