I think this is amazing. Love Kagi. I’m happy to pay for a good search.
Nobody speaks about their AI Assistant, but it is really good. They somehow harnessed it so that it mainly searches info first and sticks to the verifiable data. I prefer it to Claude or any other, because it actually answers the question I need without fluff or “Great question! Here’s some plausible nonsense you can trip over instead of doing actual research!”.
It truly is a game changer. I often use it before the regular search now. And no joke, it doesn't just save time, I've had it find things that I actually couldn't doing manual search. The reason is because it often knows when there are alternative names for the same sort of product and so can find things that are rebranded or the like.
I agree with you that I'd love to know more about how it works. The citations to real supporting websites are generally very good (though I've seen a few mistakes).
I find it to be hit-and-miss, but overall more useful than not. What I like most is its relative brevity, and the response often helps me find a better prompt to ask so that's good. It's also really fast which I appreciate.
I wouldn't say it's better, it's a great alternative though. I find that for some tasks I much prefer doing a search myself as I may not trust the sources it's pulling from, but for others it's not only adequate but much faster.
I think it's a great complement to human lead search.
"Able to find things that regular search cannot" is pretty much the value-add for AI in Google as well. The blue links are becoming less and less helpful.
I've been a subscriber to Kagi for a few years and just started using the AI assistant.
On the one hand, it's made me a believer in AI as it relates to web searching.
On the other hand, the only reason I need AI to search the web is that the web is now a impenetrable forest of near unreadable, SEO-stuffed, spammy articles.
> On the other hand, the only reason I need AI to search the web is that the web is now a impenetrable forest of near unreadable, SEO-stuffed, spammy articles.
Unfortunately, that is nothing new. A big part of the reason why Google gained popularity, and maintained it for so long, was its resistance to SEO stuffed, spammy articles. The spammers continued escalating things, and now it pretty much takes AI to avoid their garbage.
The new part is AI makes it much easier and cheaper to produce garbage than ever before. So the ratio of garbage to non-garbage has been increasing rapidly.
> So the ratio of garbage to non-garbage has been increasing rapidly.
Is it actually, really? Even late 90s most stuff was garbage, and been like that for as long as I've been on the internet, or maybe I've just always been jaded. But the ratio of "garbage to non-garbage" seems constant for as long as I can remember, multiple decades at this point. The tools for finding the signal through the noise have changed over time though.
I remember people using Yahoo and AltaVista a lot. And before gmail everyone had a yahoo email (if they didn't already have a hotmail). Times change, kings fall from grace.
Google has been the dominant force for about 80% of the internet’s existence. I understand things come and go, but Google is incredibly unique. Their importance cannot be overstated
Yes but they were mainly a force of evil. Despite their old motto. They made the internet everything it is today, a hellscape of tracking, advertising and other monetisation.
I'm not sure AI fully dodges it either. I've asked questions like "best [cuisine] in [place]" and the AI answers with a truly mediocre restaurant that has an amazing website.
Probably true. That said, I spend most of my time querying LLMs for more objective and less financially motivated information. The only time I ask for recommendations is for new routes to explore by bike, and even then it will adapt if you tell it why a route is bad.
I suspect that product and service recommendations will be a near impossible problem to solve. Any time there is a potential for financial gain, someone will figure out how to game the system. Even if it could be resolved, the queries would have to focus on objective criteria ("best" is too subjective).
But then once you get shit results, don't you iterate on the prompt and try again? In this case, add what you actually care about, not just "best cuisine" which can mean a lot of things, and also straight up tell it to not focus on the website or whatever, then eventually you'll get somewhat reasonable and applicable results.
Nope--I was on vacation, walked to the place, and saw it looked like a disaster, but then everyone was hungry and we were already out and about. Too late for more research. I picked a place the old-fashioned way, by walking around and seeing what was nearby. We did fine.
I remember at the time they were working on it a lot of people (honestly, including me) were pretty critical of them "wasting" their limited resources on AI stuff. We saw it as being sort of similar to the t-shirt drama [1].
I'm here to admit I was wrong, it's incredibly useful. It misses the mark frequently, like any LLM, but it's well annotated with footnotes that link to the original material.
I use it all the time.
I also love how easy it is to use or not - just add a question mark to the end of your query to use the AI, or leave it off for traditional search.
With that UX it doesn't feel like it's being shoved down my throat.
Do you even know what it is? Kagi Assistant is just one part of their offerings. They are a search engine first and foremost. And as you can see many people in this thread are singing the praises of all parts of the Kagi offering.
Creating an account just to hate on a company trying to shake up a monolithic and user-predatory product space is just sad.
I've configured Claude to use their search MCP instead of vanilla Websearch and it has been great. I think it's subscriber-only but I've been on the $5/mo plan for a while and have never him my limit. I might bump up to a higher plan just out of how philosophically refreshing they are as a service. https://github.com/kagisearch/kagimcp
I also confirm the $5 you get is only promotional credit.
I tested on my local LLM chat via Open WebUI, 2 requests from that cost $0.02 (so $4.98 left), and after my monthly Kagi renewal an hour ago, it did not reset, I am still on $4.98.
You are probably better off using Kagi Assistant anyways, as you get $10 there on the $10 Professional plan.
For local LLM web searches, to save on cost I will stick with the Brave Search API which gives you $5 free every month.
You know now that I think of it me too. I absolutely never used the free google one because I never trusted it (you know what they say about free things).
If I'm paying for it through Kagi, I could at least trust their incentives are aligned with mine instead of whoever is paying them the most.
> presumably they want a shopping feature so that they can shop
Right... why? Is it really too much after talking to the assistant and finding the item you want to purchase, to then click the provided link to the product page and click buy?
My assumption is something like the google shopping feature. This shows the product you're looking for at all available sites in your country so you can purchase the cheapest one.
I've been using Kagi assistant for this for quite a while. I just run a query with some variation of "Compile a list of online stores with shipping+price, sorted by the average user review score of the entire site, sub-sorted by price where I can buy $foobar?" sometimes I have to give it another push to tell it options or where to look, but this works pretty well for comparison shopping.
LLMs are not magic, nor are they gonna crawl some dozens of websites, they work with search results they are provided with, and thus they need good search results to work with. Maybe in the US or sth it is better, but where I live finding good shopping results from local shops/websites is not straightforward. Google shopping, that I just now checked, seems much better with querries that I struggled with kagi. Typically, in order to find things I need I usually run searches on local price aggregators than kagi, though it is not perfect because they do not include shops that don't want to pay them, and their search features are not great either. But running kagi for local search shopping results ime is not great, and similarly the assistant do not provide great results either.
I do not want a click-to-buy button, I want kind of lens that is specialised for shopping within specific regions (the current kagi lenses allow only 10 sites which is not really enough for shopping).
I meant, they are not gonna start from scratch by crawling websites, they are gonna rely on the search results to grab content of relevant results. I think this is what research does, in a higher quantity.
In contrast, I may start with a page in a website and start going around looking for related products etc. The assistant is not gonna start doing that (I think at least), it just gonna get the page provided by the search results. For this reason, what search results it is provided with is important.
I am also (mostly) buying online, but for certain reasons I want to buy from domestic websites rather than ones based in other countries, if possible. Kagi has not been great at providing great results for that.
I use pi and librechat and I am really limited by the lack of proper web connector. Testing Tavily right now, Kagi is next.
I wonder why I see so few comments about it on HN, does everybody have this figured out already? Am I a dummy? Am I severely overstating how important of a frature web access and search is for LLM use?
I use Claude code for basically everything, since the harness is pretty complete. The online chat interface works fine too. For basic research I downgrade the model and effort. Smarter models are good at knowing when to web search rather than guess, but you should always specify “do a web search”.
The one caveat is opus has an absolutely awful voice. The way it writes is annoying and verbose. Sonnet actually works better for this type of thing.
Yeah I pretty much soley used assistant.kagi.com the last couple months. Traditional search days are really numbered.
I'm finding though ChatGPT with sol 5.6 is as good or better, so I'm using that primarily these days. Can also do actions in browsers on your behalf, so you can have it book your whole vacation, etc. Pretty rad.
There is a lot that agents can help with wrt search, but they also have a lot of blind spots. They work well many times, but then there are times where I find stuff that multiple llms missed. Imo this is gonna work as in many other domains, as in the more different tools one is proficient with (searching with llms, searching traditionally etc) the more successfully one solves problems.
Funny how that works out, because AI is the solution to a problem significantly worsened by AI. The web didn't used to be this terrible before chatGPT, now it's just so much slop. At this point AI is required to sift through all the bullshit caused at least in part by AI.
I use it for any first pass searches. It’s great. Add a ? at the end of a search and get an annotated answer. Super happy with it and super happy to keep paying. Plus they send me stickers sometimes that go on my laptop that serve as a conversation starter to evangelize Kagi a bit in person. One of the only few services I happily pay for.
I think the first time around they were included with the free shirt they sent out. Second time they were with the free socks. Maybe general merch store orders might include them too.
Kagi's translation is absolutely top notch. I used it for a few things and while Google's, even Gemini, got some context, kagi's caught the spirit of the text perfectly.
It's that good because the language model they use is expensive I'm pretty sure. I haven't seen many details on what they use, but it's apparently a loss leader for them and already can no longer be used for free.
Meanwhile google translate is probably served on the cheapest model they could muster.
You can get it on at least iOS as well as its own app. You can select from bunch of China-models, I use Kim pretty much always. You can also do a system prompt and bunch of the same stuff you can with Claude etc. But it lies to me hardly ever. I don’t know how they did that.
I probably put !expert in half my searches and rely on Kagi for most of my research that isn't just agent context gathering (and a fair amount of what _is_ just context gathering). It's excellent and cites good sources.
Last time I compared them, it was actually ChatGPT that provided a similarly good search experience, especially compared to Claude. OpenAI doesn't meter its regular chat experience by token usage (yet), so I don't have to ration the better model and just leave it on Thinking High. The end result has definitely been better than what I got from the same queries in Claude or Gemini, and I suspect this isn't just about differences between the models, but also about the way search is integrated into their end-user chat apps.
This is honestly also one of the reasons why I don't subscribe to Kagi anymore. My 'Google-Fu', combined with whatever personalization is happening on my account, already keeps Google's search results in a similarly useful state to Kagi's, so Google and ChatGPT already cover my needs.
Paying for Kagi Ultimate feels like a privacy upgrade but a quality sidegrade that isn't worth $270 a year.
I have Kagi and ChatGPT subs and would agree ChatGPT is the better offering. But you can use Luna with Kagi Assistant on the cheap plan which works pretty well too
I became so annoyed by google's ai overview and front page results that I am using 99% duckduckgo as the least bad free alternative for the other leg from my chatgpt usage in this context.
I've also been pleasantly surprised about Brave's search based AI answers. They have been good for months and cite sources.
It will say that nothing in the provided search context has your answer which seems like a simple system prompt but somehow other models from bigger companies have a much harder time saying "I don't know" by default.
I just wish it were easier to keep up with which models are good for what. Especially with the introduction of the research assistant, I have no idea whether I should be using that or one of the standalone models. And don't even get me started on trying to pick the best open model, it's always fluctuating.
I often try 2 different models with the same prompt in case one misses anything. But most of the times I do not see much difference. Newer models do not seem to improve search that much, so I do not think you have to stress too much about new models etc. In general I stick with kimi 2.6 reasoning because it is very very cheap and does thorough searching. I do not see much gains with using more expensive models like opus etc, so I prefer to use multiple cheaper, open models simultaneously except for "harder" queries. The advantage of the cheaper open models is that you can prompt 2-3 of them and if you see the results unsatisfactory you can retry with more expensive ones. And ime sometimes models that are not talked much, like gemini pro, perform quite good.
I think there is a lot of space for improvement on "searching skills" though, sometimes the search strings they use are not great.
> They somehow harnessed it so that it mainly searches info first and sticks to the verifiable data.
I like this too, but it isn't as bulletproof as it sounds. I've run into two instances in the past week where it confidently gave me wrong information, because it ran into a search result that had wrong information, and it trusted that information.
You could perhaps say that's fine; if you were doing the search yourself, you'd get the same results and might trust the wrong information too. But the "might" there is... oh boy, I'm gonna say it... load-bearing: someone is much more likely to believe the confident LLM response than if they were looking directly at the primary source, along with search results with other sources that might contradict it.
Yes, I use AI assistant a lot. I copy the agent definition from Geir Isene to a custom assistant, it works nicely. In my experience, Kimi K2.6 reasoning works best.
Another +1 for their Asssitant from me. I have the app on my iPhone, and despite having Perplexity/OpenAI/Claude/Gemini paid plans I find myself gravitating to Kagi for q&a.
Is it actually good, like 2006 good? Can I go to page 300 rabbit hole and find some obscure website nobody knows with interesting content on the topic I am researching?
Unfortunately, those days might be good no matter what - a lot of those sites have disappeared with their hosting platforms (e.g Geocities). That said, I have been able to find them on Kagi with a bit of effort.
Kagi also has a Small Web feature which is meant to surface exactly that sort of thing. It's not amazing, but I appreciate it existing. I think you can get a "random small web page" kind of thing even without a subscription here: https://kagi.com/smallweb/
Kimi K2.7 Code, for me. It's the first model I've used that usually writes working code, somewhat rarely breaks stuff when adding or changing features, and usually fixes a bug correctly just by providing a traceback or description.
Kagi Assistant defaults to the last-used model unless you got there from a Quick Answer. Several times I forgot to change models, and interrupted K2.7 Code, but after a failing to notice the model a few times and getting usable results, I now usually just leave it alone. Unless I want a quicker answer and doubt a "Code" model will grok me.
> > "Nobody speaks about their AI Assistant, but it is really good. "
So this has nothing to do with the Kagi AI assistant then... it's just the model that's perceived as "good". My point is, there are cheaper ways to get access to the model, without a Kagi subscription...
I was subscribed to Kagi for a while but cancelled because there was no subscription for JUST search that didn't include their inferior AI assistant. Why am I paying for it?
If you compare the Professional plan (unlimited searches), to the API, you could justify that the extra AI cost is negligable if you are heavy user. The API costs $12 per 1000 searches, so 1.2¢ per individual search. 10/0.0012≈833.33. At 834 searches or more you would not be paying for the AI features. Obviously, this is only a comparison to the API, which might not have all special Kagi features.
But if you're using the starter plan, which is 300 searches for $5, then you definitely do pay for the assistant. The API cost would be $3.60.
Nobody speaks about their AI Assistant, but it is really good. They somehow harnessed it so that it mainly searches info first and sticks to the verifiable data. I prefer it to Claude or any other, because it actually answers the question I need without fluff or “Great question! Here’s some plausible nonsense you can trip over instead of doing actual research!”.