I'll be direct in this post — not to be combative, but because I think honest signal from paying customers is valuable. I hope it reads that way too! And sorry it got so long.
Background
The reason I am here is I want the best search experience possible, and I want reasonable AI integration where it makes sense.
The recent changes that Google made where search essentially became a chat experience by default would have made me really unhappy. So I am actually happy to be here. I like the control Kagi gives me — requesting an AI summary, choosing the model and, when it makes sense, choosing to opt into a chat experience with a range of models and agents. I also trust you to choose vendors that promise not to use my chat as training data.
But I don't see you doubling down on this, and I think you should. The chat and search experience is your main lever for increasing subscription value. I appreciate the efforts to make Kagi useful to as many people as possible and as independent from Google as possible — Orion, translate, news, email — but not at the cost of neglecting the core value proposition. Frankly, none of those features move the needle for me personally, and I suspect many power users feel similar.
Today's Kagi Assistant experience
Kagi Assistant isn't bad, but it's lacking compared to other chat experiences, and it's an island I can't connect to anything else. It feels static — like the vision is that it will never do anything beyond (re)searching the public web for me.
A quick detour into missing features that directly affect the economics of using Assistant. I usually didn't come close to maxing out my budget until this month, when I started using Claude Opus 4.6 heavily as a sparring partner for personal finance and real estate questions. The reason it maxed out? Kagi Assistant lacks features that would have saved significant tokens:
- Memory. I wouldn't like it across all threads by default, but project-specific memory that carries into new threads would be useful, so I can avoid adding onto long conversations where token cost grows exponentially with each reply.
- Summaries. Many agents can summarize history (e.g. using a cheaper model, with the context of the latest request). This could be opt-in.
- Sub-agents. Assistant already has some sub-agent functionality (I notice it queries one for large file uploads) — but I can't instruct it to dive into a topic with a sub-agent and return just the summary.
- Context/cost visualization. There's a notice near the limit, but some other agent systems visualize context depth continuously, which helps decide when to open a new thread.
Small UX note: I have ADHD and regularly forget to click the "temporary" tag to persist a thread. This costs me real time.
Where the API comes in
The chat assistant and agent space is fast-moving. Many of your power users like to build in this space, often with private sensitive information, and they tend to be very privacy-conscious. So why is Assistant so closed off? Even Kagi MCP has no included API fair use, not even a single free search request from a custom agent implementation?
I'd love a Kagi Assistant experience where I can store context locally on my machine, customize the system prompt, connect MCP servers, manage memory, and orchestrate sub-agents. The building blocks are largely there — you've prototyped a CLI as a bash script, you have the search MCP — but there's no way for users to assemble them.
To be clear about what the API isn't for me: a cheap coding assistant. My employer provides those, and Kagi isn't an option I'm allowed to use at work anyway. In my free time I code occasionally and it'd be nice, but daily use would blow past the $20 limit quickly. I don't think most people come to Kagi for that.
What I'd like to see
I just want a good Assistant that is truly useful and works with my favorite search engine. While the Assistant isn't where it should be, offering some API fair use would be — fair.
Concretely: Let users allocate their Assistant and API budget as they see fit, within a window that allows you to still make money. Give them the option to buy additional tokens if they need them. But please don't make the API pay-per-use only.
I think it might be smart to double down on Assistant by offering a more advanced, local-first version — a Tauri-based desktop app with memory, MCP support, and a CLI — possibly open-source. That might be a genuine differentiator.
