Choosing the right tool for the job

There are hundreds of good, self-hosted tools for any one job — and almost none of them wear a label saying "pick me." The hard part of building your own stack isn't the deploying. It's the choosing. One wrong pick and you're six months into a tool that wants a different database than you gave it.

Here's the system I use, plus the head-to-heads I found myself running. It's not exhaustive — it's a way of thinking that works.

The method: decide the job first

Before comparing tools, I write one sentence: what am I actually trying to get done? Every tool gets judged against that, not against a feature list. "Chat with my documents" and "run a full AI assistant for my team" look similar and are completely different jobs — they lead to different picks.

Head-to-head: chat with your own documents

If I just want to talk to my files quickly, AnythingLLM. If the documents are messy — PDFs with odd tables and layouts — RAGflow earns its keep. If the whole point is that nothing leaves my machine, PrivateGPT.

Head-to-head: build an agent without being a developer

Flowise gets you an agent fastest. Langflow if you want more surface area. But n8n is the one I'd install if I only installed one thing — it's not just agents, it's the glue for everything else.

The rule that saves me

Pick one winner per job. The dashboard you open twice a year is maintenance debt, not infrastructure.

Every tool you run is a service you maintain and a port you keep warm. The right stack isn't the biggest one — it's the smallest one that does its job well.

This fits into the bigger question of whether you should own your stack at all — and once you own it, what to do when the tool starts doing the work.