Baidu Comate

You will judge this on the agent, not the completion. The pitch is that you describe a requirement and it implements the thing end to end, including standing up whatever it needs to run, and the deciding daily trait is how much of that you trust on a codebase you did not write. Next-edit prediction is the quieter feature and probably the one you will use more often.

Pick it if your work already sits inside this vendor's cloud and tooling. Pick Tencent CodeBuddy when you want a general assistant with a wider surface and less ambition per prompt.

Read that capability twice. An agent that configures an environment and launches services on your behalf is executing arbitrary setup on your machine, and the row records no container isolation of any kind, so the blast radius is your working tree and whatever credentials sit beside it. The convenience being sold is precisely the removal of the moment where you would have read the command first.

What it does right: next-edit prediction is a small, checkable feature that fails visibly and costs nothing when it is wrong, which is more than can be said for the headline.

1. Context gathering is the central claim and the least documented part: understanding the business logic of an existing codebase implies an index, a ranking and a budget, and none of the three is described anywhere in the published material. 2. Edit application across multiple files is stated without a diff format or a review step. 3. Verification is not mentioned.

No benchmark is offered, and the only English-language source is a marketplace listing. The observation: the harder the claim, the thinner the documentation supporting it, which is a pattern rather than an accident on this row.

This is the vertically integrated case, and it is the strongest structural position in the category. Owning the model means the cost of goods is infrastructure rather than another lab's invoice, so the price can fall without the margin following it. The ladder from individual editions to dedicated and on-premises deployments is a well-built upsell path aimed at exactly the buyers who pay most.

There is no acquirer, because the parent is one. Position: durable inside its home market, structurally uninteresting outside it, and the tool's fate is tied to the platform strategy rather than the product.

The extension itself is listed free, which gets it onto laptops without a purchase order and is exactly why I want it stopped at the gate instead. Every paid tier above that is quoted in a currency our finance system does not carry a rate card for, and the enterprise and dedicated editions publish no figure at all. Identity integration is undocumented, retention is undocumented, and nothing here reaches our pipelines.

Onboarding is an hour. Not yet, and the blocker is disclosure rather than capability.

Every lever I look for is missing. The backbone is a single family of models I cannot substitute, there is no way to supply my own credentials, nothing runs on my own hardware, and there is no client for the tool servers I already maintain. The extension is closed, so I cannot read what it sends or when.

I will give it one thing: the vendor never dresses a hosted service up as an open platform. That is not a compliment I can convert into anything useful. This is a rental with no keys, and I would not point it at a repository I care about.