Switched to Codeium for the privacy, stayed for the price.

That "shockingly context-aware" chat is the real deal for navigating terraform modules. It gets the dependency chain in a way others just guess at.

You nailed the local model being the killer feature. It's not just privacy, it's predictability. No surprise latency spikes because some cloud provider's region is having a bad day.

But watch the model version you lock into with that offline setup. If you bake it into a team image and forget it, you're stuck with old bugs and patterns while the hosted service improves. You need a rebuild pipeline for the container, which adds ops overhead.

Great question. It's more of the first one - the suggestions become less relevant, or it starts re-suggesting patterns you already corrected a few prompts back. I haven't noticed a significant performance dip; it just feels like its "mental map" of the project gets full.

For your analytics scripts, you'll likely hit that wall when you start abstracting common functions into separate utility modules. It handles a linear chain well, but struggles with the cross-talk. My workaround is to restart the chat and give it a fresh, one-sentence summary of the architecture when switching focus to a new module. It's a minor friction, but keeps it useful.

> Local model option. This is the killer feature.

Only if you actually deploy it that way. Using their cloud negates the core privacy claim. Their terms still apply.

The price isn't suspicious, it's a commodity price for an inference engine. The incumbents are overcharging for the wrapper. Your real savings is skipping the vendor security review.