This product has an unusual history that explains most of what it is now. It was doing AI code completion before the current wave, back when that meant a smaller model trained on public code producing single line suggestions, and it was good at that. Then the field moved, the frontier models arrived, and a set of competitors appeared with vastly more capable underlying models and much larger budgets. Rather than trying to win on raw capability, which was not winnable, this product turned towards the segment that could not use the competitors at all.
That was a rational decision and it is the reason to consider it. The compliance position is real and it is not marketing. There are organisations, in defence, in regulated finance, in healthcare, in parts of government, where sending source code to a third party API is not a policy preference but a prohibition. For those organisations, most of the market simply does not exist.
This runs on your own infrastructure, including fully disconnected from the internet, which almost nobody else offers seriously. If that is your constraint, the evaluation is short, because the field of candidates is small. The provenance work is the other genuine contribution and it deserves more attention than it gets. There is an unresolved legal question about code generated by models trained on licensed code, and most vendors handle it with an indemnity clause and a change of subject.
This product will tell you when a suggestion closely matches known public code and what licence that code carries. Whether that fully addresses the problem is arguable. It is more than a shrug, and for a legal team trying to form a position it is something concrete to work with. Not retaining your code as the default rather than as a setting buried in an admin panel is a small thing that reflects the same orientation.
Now the honest assessment of the actual experience, because that is what most readers care about. The completions are behind. Not unusable, and behind in a way you notice within a day if you have used the leaders. Longer completions are less coherent.
The understanding of the surrounding codebase is shallower. The multi-file and agentic capabilities, which is where the competition has moved decisively, are considerably less developed. If your reason for adopting an assistant is to write code faster, you would be choosing a worse tool for that purpose, and no amount of privacy compensates for it unless privacy is the binding constraint. As a learning resource, which is how I am assessing it here, it is weak.
The material on the site is sales content. There are feature pages, comparison pages that unsurprisingly favour the author, and case studies. There is very little that teaches you anything about how to work well with an AI assistant, what these tools are good and bad at, how to review generated code sceptically, or how to tell when the model has confidently produced something wrong. Someone learning the craft of working alongside these tools will get more from a good blog post than from anything published here.
Pricing is the last consideration and it only resolves one way. Per seat it is roughly comparable to the competition, and the competition is better at the core job. The purchase makes sense when the deployment model is a requirement rather than a preference, and it does not otherwise. My three point two is for a product that genuinely serves an underserved and constrained market, with a self-hosting story nobody else matches and a serious attempt at the licence provenance problem.
Marked down for suggestion quality that has fallen behind, for agentic capabilities that lag badly, and for learning material that is entirely promotional. If your security team has already ruled out the alternatives, this is your answer. If they have not, buy something else.