okay this video is going to be an interesting one we're going to talk about the new tavali tavali tavilla tavil tavali I think it's pronounced tavil we're going to talk about the new tavil mCP server first I'm going to give you some context so tavil released their own official mCP server and until now there was a MCB server created by the community T has been around for a while it is a search and scraping tool made specifically for llms I found it after using serper for a few months I just found when I used tavil over serper I got better results for what I was looking for and I tried the community mCP server for T and I didn't find it that great so I went back to Brave church and brave search has been awesome and I've done a lot of videos on using Brave search but now that tavil finally released their own mCP server I wanted to integrate it the difference between Brave and tavil is brave is really good at doing broad internet searches it's been really useful when I combine it with sequential thinking with Fetch with Puppeteer inside claw and with cursor but sometimes you want a little bit deeper research you want to go in there and get the details and while I was able to get this with brave eventually T usually gets to the point really quickly so let's just switch gears really quickly and you'll see where I'm going open AI over the last 2 months has released a lot 01 full they release Sora they release a few agentic things like tasks operator and deep research tasks to be honest has been a bit of a let down I keep trying to make it work for me and it doesn't do exactly what I want I haven't found operator to be as lifechanging as I thought it would be at least not yet there's a lot of limitations there and I'm still playing with it and once I find something that I think think is actually valuable I'll make a video on it but we're not there yet but deep research by open AI is amazing I've been using it daily and I find it to be so powerful I don't want to overuse this analogy but I do think that open ai's deep research really is that jump from gpt2 to even GPD 40 just the detail of the research it does and then the detail of the paper that comes out is amazing and I found some really cool things I'm not going to share it all with you I'm going to share one thing with you and then how I paired it with notebook LMS so I've spoken about uh my fiance we we just came back from treatment abroad and we're just trying to figure out you know how to boost our energy in a safe way and one thing that came up was creatine and I don't know anything about creatine what I ended up doing was asking deep research to look into creatine for me and for her to understand is it safe for her to take in her situation as immunosuppressed came back 6 minutes later with a really detailed response the pros the cons for someone that is recovering from cancer this is just one example of some of the deep research I did and I've actually been factchecking with various sources to make sure this is real this is not a hallucination and that's hard in and of itself but so far I found that everything I've put into deep research has come back like 90% correct so I'm very confident that deep research while it might have hallucinations is much lower than GPT 40 or any modern llm right now because it really goes through the iterative process and kind of weeds out the so anyways what we get is this really long paper and it's great but it could take you a long time to read it I had an idea of taking this whole thing and then having notebook LM explain it back to me and it is such a cool workflow so I'll just show it to you so I took this whole this whole output that we're still scrolling through and I put it into notebook LM you can just have it create a podcast based on all the sources you put in so in short what I've been doing is I've been using open ai's deep research putting it into notebook LM and having it create a podcast and then just listening to it and because with deep research from open AI we get a PhD level report putting that into notebook LM and then listening to a podcast is amazing it's I've just learned so much from this and this is a powerful workflow that being said it is not a cheap workflow so I've seen a lot of people replicating deep research or trying to replicate deep research I've tried some of them out I haven't seen anything really replicate open AI deep research in quality that being said I'm sure we're going to get to a point where there will be a full open source deep research that is comparable to open AI tool so anyways released their mCP server their official mCP server and around the same time I saw this post that they made on link in about do yourself deep research so it's this whole iterative process that uses T to create a deep research type response I haven't tried this one yet but I started thinking can we replicate this on some level with claw desktop and mCP servers so that's what I've been doing so first I'm going to show you how to add tavil to mCP and then I'm going to show you what I've been building and it's not perfect yet I've been iterating on this for about 24 hours now I keep hitting those rate limits having to wait and again this is not building a whole thing like they did here because this is all within claw desktop meaning the whole way we're doing this is with custom instructions which in and of itself can be tricky the more instructions you give it but I think it's been getting better I want to show it to you just to show you tavil has a free plan just like Brave you get about 1,000 API credits per month to compare that brave search gives you 2,000 API credits per month so what I've done is I start my deep research with brave and then I go into T and that way I'm able to take a little credits from here a little bit credits from here and not waste my T credits and what's really cool about t they also have this pay you go model here is T's own mCP server you can see it's hosted from their own GitHub they have two main tools within the tavil mCP server there's search and extract and using them together is really powerful this should they install really quickly same as always if you're doing this with Claude you just going into the MCV servers you just take this part down to here copy and paste it into your config file and by the way some people have written to me about installing via smithery if you haven't noticed I like to have a level of control so this is how I like to do it regardless okay so now we're in vs code and as you see here I just went to the bottom of my config file added a comma and I added this in so you paste this in and of course don't forget to add in your API key here you save your config file and commit it to get now back to Claud now I've created this new project called deepest research toil plus sequential thinking I just going to show you my custom instructions so I've been iterating on this a lot I added XML added so much to this and I think it's doing pretty well in this project we're going to use Brave search T and sequential thinking in earlier iterations I also included Fetch and Puppeteer and a bunch of other tools but I realized that by simplifying it just to these three main tools I still get a great output without confusing CLA so I wrote this whole custom instructions I'll share with you the bottom line This Cloud project uses sequential thinking Brave search and T to do deep research now it's not giving me a very long response like open AI deep research and that's that's for various reasons from a fundamental level within CLA desktop we're not going to be able to replicate open ai's deep research one thing I took from Deep research from open AI is I baked in the clarifying questions and I also took what I liked from Deep research from Google and I added a step to verify and confirm the research plan and I found that being able to first of all focus in by clarifying questions and then being able to redirect Claude by looking at its plan and saying no change it before it goes into research gives me great responses I'm just going to show to you really quickly I'm going to put a similar prompt in what are the benefits of taking creatine for a 35-year-old male that works out three times a week I put these like style points in so clarifying questions no pre-existing conditions well only baldness okay so I answered it's clarifying questions now step two is giving me the research planning and it will only move forward if I confirm it it's also planning out what tool is going to use for what part so you see here Brave search for broad studies tavil for specific performance metrics tavil for detailed neurological studies sequential thinking for pattern analysis so here's a research execution plan begin with brave search use sequential thinking deploy toil search integrated fightings across themes using sequential thinking focus on practical applications so I'm going say yes so you see here it's really going through a brave sequential sequential sequential sequential sequential Lilia sequential Brave sequential sequential sequential so now it's doing the final report took a lot quicker than deep research by open AI the quantity of output is also a lot shorter than open AIS but but it's detailed and I think it comes out pretty good that's funny it didn't get my joke there it took my baldness thing as something seriously multiple subsequent Studies have failed this demonstrate direct cation with hair loss okay so that's what it came up with again it's not perfect but I think it's pretty cool I don't think we're going to be able to replicate open ai's deep research for various reasons but what I like about this is a bit shorter than deep research I definitely suggest putting mCP to the side and trying open eyes deep research if someone has an interesting idea for something I should put into deep research tell me I'll run it in open a deep research and then I'll put into notebook LM and I'll show you cuz it's really really cool let me know if you think T is better than Brave search let me know if you want to try my deepest research Cloud project I'm going to share the custom instructions anyways I hope you found this video helpful or insightful if you have any feedback or questions drop them in the comments below especially if you have an idea for what I should test with deep research either with open AI or with deepest research if you have any ideas on how to improve the custom instructions here let me know if you haven't done it already don't forget to subscribe to this Channel and also like the video it really helps me grow thank you for watching and have a great day