aiXcoder

You will consider this for one reason, which is that it deploys on infrastructure you control rather than somebody's cloud. Everything else on offer, plugins for two editor families and a terminal companion, exists elsewhere. The deciding daily trait is that the perimeter question has an answer, and for the teams who need that answer nothing else on this board is a substitute.

Pick it if a sovereign deployment is the requirement and your procurement runs through China. Pick Tabnine when you want the same isolation with a vendor your legal team has already heard of.

The problem is evidence. A suite spanning requirement analysis, code generation, testing, wiki and a terminal companion is a large surface, and the whole of it is described by a single marketing page with no repository, no release notes and no independent write-up behind it. A tool you cannot inspect is a tool whose failure modes you learn about after the contract is signed, at your own expense.

What it does right: the plugin exposes distinct Agent, Plan and Ask modes, so the level of autonomy is a deliberate choice rather than an assumption.

1. The division of labour is sound in principle: a large model plans and a small specialised model applies the change, which is where most token spend hides and where a narrow model can be trained to beat a general one. 2. That applier is published openly at four billion parameters, so the claim is at least partially checkable. 3. Verification is delegated to a separate testing agent, though the handoff between them is undescribed.

No benchmark is offered for any of it. The observation: the interesting engineering is real and the documentation is one page long.

Read the delivery model and the company becomes legible. Sold by seat, negotiated per contract, installed on the buyer's own compute: that is high revenue per logo, slow growth, and margins set by how many engineers each deployment consumes. Pricing power is genuinely strong, because a sovereign deployment has few substitutes and the buyer cannot easily leave once the stack is inside their perimeter.

Likely acquirer: a domestic cloud vendor that wants the developer surface above its own inference. Position: a durable regional business, not a category winner, and not investable at a Western multiple.

The demo covers a lot of ground and I cannot budget any of it. There is no list price, no free tier and no trial I can start without a conversation, which means the first meeting is with their sales team rather than my own engineers. The on-premises deployment is the part security would actually like, and it also means we take on hosting, capacity and upgrade work that a hosted tool absorbs.

Onboarding cost is unknown for the same reason as everything else. Not yet.

The thing I cannot argue with is where it runs. This is served from compute the customer controls, so inference never leaves the building, which is more than every hosted assistant on this board can say. It calls MCP tools, so my own servers are reachable from inside it.

Everything else is a wall. The source is closed, I cannot supply my own key, and the model selection is theirs. Sovereignty over the hardware without sovereignty over the code is a rental with a nicer address, and if the vendor stops shipping I have nothing to fork.