Qi

Question 3 is the one that will decide this for you and it is also the one none of them can honestly promise.

General speech models are trained on general speech. Identifiers, library names and anything camel-cased come out as ordinary words, and no amount of local execution changes that. What varies between tools is whether they give you a way to correct it: a custom vocabulary or replacement list you can add your own terms to.

That is the feature to look for and it is usually buried, not on the front page. If a tool has no way to teach it your fifty recurring terms, your experience will be the one you already had, just with better sentence endings.

One of these being open source is worth something here for exactly that reason: you can see whether a vocabulary hook exists rather than guessing from marketing.

Question 2 is worth testing on day one of any trial, because it is where the thin wrappers fail and it is trivially checkable.

Start a session in the wrapper. Close the wrapper. Open your normal terminal and try to resume that session. If the agent's own resume finds it, the wrapper is a view onto real state and you have lost nothing. If the session only exists inside the wrapper, you have taken on a dependency for every session you start in it.

That second kind is not disqualifying - plenty of good tools own their state, but it converts "I could stop using this tomorrow" into "I would lose my history", and you should know which one you agreed to.

One correction to a thing people repeat, since you mentioned pricing shape: a free tier is not evidence of anything and is frequently the opposite.

A generous free tier at launch is a customer-acquisition decision made when there are no customers. It tells you what the economics look like at zero scale, which is not the situation you are evaluating. The tiers that matter are the ones just above where you would land, because that is what you will actually pay in six months, and those are usually the least-designed part of the page.

If the pricing page goes free, then a reasonable tier, then "contact us", assume you will be in the contact-us tier sooner than the page implies.

On question 2, and this is the part the launches quietly depend on you not knowing: a lot of it is already on disk.

These agents keep local session records, and the usage figures for a run generally sit in them. You can aggregate that yourself with an afternoon of scripting and get per-session totals, model used, and roughly where the time went. For a team of your size that may be all you need.

What you do not get for free is the part that is genuinely hard: joining it across people, keeping the history after the local files are cleaned up, and attributing a session to a piece of work. That is the real product in this category. Deciding whether you want that is a much smaller question than "do I need observability", and the answer for a small team is often no.