CodeGeeX

The trait that decides it is the inline chat on Ctrl+I. You highlight code, ask, and get an answer in place, which covers explanation, a unit test or a small fix without leaving the file. For a student, a hobbyist, or anyone whose employer will not approve a licence, that is a lot of function for no money.

What it will not do is carry a task across files. Pick Cline when you want an agent that plans an edit, touches five files and shows you the diff, and keep this as the free assistant it is.

The architecture has one exit and no valve. Completions are generated remotely, there is no self-hosted path and no option to substitute an endpoint, which means the code in an open editor travels to a service chosen by the vendor. For anyone under a contract that restricts where source may go, that ends the evaluation regardless of quality.

Treat it as a public service and act accordingly with proprietary work. What it does right: it names the models behind it rather than hiding the whole thing behind an unnamed assistant, which is more disclosure than several paid competitors offer.

The capability set is deliberately bounded. Context reaches the model two ways: the surrounding buffer for completion, and retrieval over a repository index for questions about a wider codebase. Actions are suggestions the developer accepts, so there is no autonomous application of edits, and consequently nothing that could apply a change wrongly while nobody is watching.

That bound is why the architecture is uninteresting and honest at once. It generates text; a human is the verification stage. The underlying model has a published research lineage as an open multilingual code model, and no benchmark is recorded here to compare against it.

Free forever is not charity, it is customer acquisition for the model underneath. More than 1.3 million installs of a single listing puts the vendor's model family into daily contact with an enormous developer population, and habit is the cheapest moat available in this category. The plugin is the sampling table; the model business is the shop.

Pricing power today is zero by construction, which is the point. Likely path: monetisation shifts to the model platform, not the extension. Position: the giveaway is rational, and the interesting asset is one layer down from what you are installing.

Zero cost for sixty seats is the shortest budget line I will see this year, and it is not the number that decides this. There is no single sign-on, no administrative console, no audit of who installed what, and no documented data-retention commitment, which means I cannot answer the three questions our customers ask us about subprocessors.

There is also no unattended mode, so it contributes nothing to our pipelines. Onboarding is a marketplace install, which is precisely the problem, since it needs no approval to spread. Not yet, and an extension policy in the meantime.

This is the split I object to on principle. There is an Apache-2.0 repository with a real open code model in it, and then there is the extension everyone installs, which is closed and talks to a service. The good open artefact and the useful daily tool are not the same thing, and only one of them is mine.

No key of my own, no local endpoint, no MCP. If I want the model I can go get the weights and serve them myself, and then I am writing my own editor integration. Some evening I might. That option is worth something, and it is not the same as ownership.