In Kagi Assistant, I would like to see the ability to be able to add and use Ollama models. Ollama has a well-documented API that would allow Kagi Assistant to communicate with both locally running Ollama instances, as well as remote ones. If the used Ollama instances are controlled by the user, this would add another layer of privacy. This feature would also greatly expand the models users are able to use in Kagi Assistant. Models like Gemma by Google, Phi by Microsoft, the countless other models available on Ollama, or even models the user has trained themselves.
As explained above already, users could use this to add another layer of privacy to models that Kagi Assistant already offers, like Llama and DeepSeek models. Users would have access to dozens more models that have already been created, trained or distilled by others, or use their own models. It would bring all of these models together in one place, which they are likely already using for the models they get access to through Kagi. To be perfectly clear, I am not a paying Kagi user (yet). Kagi looks really interesting to me and I am on the fence about it. So far my pros and cons list consists of a hell of a lot of pros, and only one con. This would add yet another pro to the list, and it might actually push me over the fence.
There are many browsers or apps that already have this feature, or similar features. Examples are Brave Browser's Leo, which I currently use. There's Continue for VSCode which allows for chat/prompting, as well as code autocompletion using both proprietary and self-hosted Ollama models. As a final example there is Open WebUI, whose Ollama-features extends beyond prompting and generation and allows you to manage downloaded models in Ollama. Many more frontends that are able to interact with Ollama can be found on their GitHub README. Open WebUI is the frontend that is most similar to Kagi Assistant. Leo functions as both a browser sidebar, and a "full page" (brave://leo-ai). Continue has a sidebar in VSCode, as well as the previously mentioned code-completion in the editor itself.
