I released 0.6.0 of koverGate and 0.3.0 of Knosh over the weekend.

koverGate 0.6.0 adds its own exclude {} configuration block in Gradle, with near-identical syntax to Kover's exclude {}. The only thing that koverGate does not support is exclude-by-annotation (annotatedBy()). Among other things, this lets you leave Kover alone from an exclusion-rule standpoint and just filter the koverGate reports, if you so choose.

Knosh 0.3.0 is a bit larger. The highlights:

• If you run knosh in a terminal, you get a status line, with one of those cute little Braille-dots animations to show a hint of progress.

• You can define defaultAgentId in the configuration JSON. If set, all commands use that agent unless the command-line specifies otherwise.

• prompt is now the default knosh command. Combined with the above defaultAgentId, running a prompt is literally as simple as knosh "this is where you tell the LLM what you want done".

Knosh 0.3.0 also has a bunch of security fixes, hinted at in a previous issue.

Simon Willison pointed out the release of Nativ. This is Ollama-ish, but macOS-specific. It focuses on MLX (LLM-on-Metal) models and is written by the developer of MLX-VLM, a key Python library in the MLX space.

The thing that I like about Nativ is that is has per-prompt and overall analytics, in terms of tokens, processing speed, RAM usage by the model, etc. It also seems to have a better UI than does Ollama with respect to model management.

I tend to run my local models on a separate machine from my development machine, so that GUI is of modest use. Nativ does offer a server mode, akin to Ollama, though it is unclear what advantages Nativ might have over Ollama in this area.

It's something that I'll be keeping my eye on, and if I start to use it, I'll be sure to post about it here.