Coding agents are quietly shifting from "pick our model, use our cloud" to "bring any model, run it yourself" and it feels like a real inflection

Been noticing a pattern across the newer AI coding tools and wanted to see if others see it too.

The first wave (Cursor, Copilot, Claude Code) all share the same shape: the tool is tied to a model or a small curated set, a lot of it runs through the vendor's cloud, and you're basically renting into one company's stack. That was fine when only a few models were any good.

But now that there are a dozen genuinely capable models, and strong local ones via Ollama/LM Studio , that lock-in is starting to feel outdated. And a new crop of tools is being built around the opposite assumption.

The clearest example I've hit is Zero (open source, github.com/gitlawb/zero). The whole pitch is "your model, your machine, your rules" — it talks to 24+ providers, you can switch models mid-task, it runs locally, and it stores nothing remotely (no telemetry). The model is a swappable part, not the identity of the tool.

What's interesting to me isn't the specific tool, it's the architectural bet: that inference is becoming a commodity you route to, the way we already treat storage or compute. If that's right, "which model does your coding agent use" becomes as weird a question as "which brand of electricity powers your laptop."

Do you think provider-agnostic, local-first agents are actually the future here, or does the convenience of an all-in-one cloud tool (Cursor etc.) win for most people regardless? Curious where people land.