CodeGPT

You will appreciate this most if you have already been burned by a subscription meter. The model choice is yours, so what you spend is what your provider charges and nothing is marked up on the way through, and that predictability is the deciding daily trait, because the thing that makes people abandon an assistant is not a bad suggestion, it is an invoice they cannot explain.

Pick it if you want that control with a polished extension around it. Pick Cline if you want the same arrangement and would rather read the code that holds your key.

The trust boundary is in the wrong place. Bringing your own key sounds like control until you notice the key now lives inside a proprietary extension, sent wherever that extension decides, with no isolation recorded anywhere on this row and no way to audit the traffic short of a proxy you set up yourself. A credential that can be spent is exactly the thing you should not hand to an opaque process.

What it does right: nothing is applied without review, so a wrong proposal costs you a keystroke rather than a file.

1. The documented order of operations is sound: plan, then read files, then propose, which puts the cheap reasoning before the expensive editing. 2. Tool-protocol connections extend context to external documents and databases, so retrieval is delegated to systems that already know their own contents rather than to an embedding of them.

What is missing is identification. The economy models offered on the free tier are unnamed, so two users comparing results are not comparing the same system, and no evaluation of the plan step is published. The observation: a product whose differentiator is planning has published nothing about how well it plans.

The asset is the install base and nothing else, because a product built around the customer's own keys owns no inference margin and no data. Two million reported installs is genuine reach, and ten dollars a month is what you charge when the thing you sell is convenience rather than capability. That is a real business and a fragile one, since the editor vendors keep absorbing exactly this layer.

Likely acquirer: a model provider that wants a distribution surface in two editor ecosystems. Position: small, and read the install number as reach rather than revenue.

One sentence on the demo: it proposed a sensible change and waited. The arithmetic is $1,800 a month at the monthly team rate, which is defensible, and then the operational problem arrives. This model of working assumes each engineer supplies credentials, so we either issue sixty scoped keys and rotate them, or we discover during an incident that somebody used a personal one. No identity integration is documented, and nothing runs unattended for us.

Onboarding is an hour. Approved with conditions: centrally issued keys, rotation on a schedule.

I will admit this one earns its place. Two different local serving options are named outright, so the whole loop runs on my hardware with nothing leaving the machine, and the provider list is long enough that switching is a dropdown rather than a migration. The tool-protocol client means my existing servers attach without a shim.

The complaint is unchanged and unfixable: the extension is closed, so I am trusting a binary with my endpoint and my key, and if the vendor disappears there is nothing to fork. Everything good here is a configuration surface. None of it is mine.