NEAR AI Ecosystem - What Did You Ship This Week? #19

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:16

   Near AI, what did you ship this week? Hello. Hello everybody. Uh, this is episode 19 of Near AI and your weeks. What did you ship this week? My name is Cameron Dennis. Uh this is essentially a distributed uh engineering standup where people in the ecosystem share within the building. Uh today is a relatively smaller agenda. Should only take about 30 minutes or so. Uh where we have uh Harsh to talk about uh his contributions

2. 0:45

   to the near MCP network and sort of the work that's being done there. Micro from Bitta uh talking about Coinbase integrations, XMTP uh building a boiler plate uh for others to launch similar agents. And last but not least, Alex Newman from Human Protocol. Uh without further ado, we'll love to kick this thing off. Uh Harsh, are you ready to come up?

3. 1:07

   So turn on your camera and I'll add you to stage. Hello. Welcome. Welcome. Welcome back. Yeah. Hi Cameron. Uh so yeah last time when I came on near AI uh live stream I think it was roughly around a month back stream 14 I guess right so at the time I presented bitcoin agent um so it's an AI agent that runs inside bit interface that gives gives you access to basically send transactions on bitcoin l1 using your near account and all of that was

4. 1:36

   possible because near has something called near uh chain signatures right which uses multi-party computation to sign these multi chain transactions in this decentralized manner and the current infra that exists on like bit which we use for Bitcoin agent as well really completes the stack to build and embed this agent and the best part is that tools that you build for one agent

5. 2:00

   are composible and can be used by other agents to do the same. So why I'm telling you all of this is because uh we are using the same knowledge to expand near MCP right. So the beauty of model context protocol is that uh that you you can basically leverage all of this tooling and make it compatible with uh any AI agent that you have. So as I told you MCP tools are composible and beta team also did launch this uh with all of their agents where you where every agent is now compatible with MCP and you get

6. 2:34

   MCP link with every agent and this is what makes it usable right. So Coinbase also yesterday did u this where there was a direct integration of MCP compliant agent. So MCP is definitely the standard. So the whole point what I am telling you is that MCP is definitely the standard that world is adopting for AI agents. ignore the nays because recently uh founder of vers guly rosh

7. 2:58

   also did this satarical post quoting lionus towells right so he told that lionus did his hobby project which is Linux uh with caption that MCP won't be big and professional and it won't be used by hobbies right so we can just go about our day and keep building with MCP uh so let me just share uh what we built right so I'll just share a couple of things here that we were able to do with near MCP server codebase. Thanks to Alex, by the way, he's the near MCP server developer. And the server had a

8. 3:28

   lot of tools already related to account and access key management, token operations, and other smart contract uh smart contract interactions. We expanded on this work and we built the chain abstraction tooling. We had the tooling for Bitcoin agent as I shared last time.

9. 3:42

   Uh we I shared the work on Bitcoin agent. So, we imported it into near MCP server and tested it. And here are some of the results. Quick heads up uh server is still under development. It is available for developers. So if you are a developer, you can use this tools today to build agents on top. Here I will show you how you can add this in claw desktop which is one of the famous MCP client right now which is the biggest and then you can get access to all of these tools and basically test

10. 4:07

    them. There will be some advanced steps like adding local key pairs to sign onchain transactions and pay for gas. So we're going to skip them because this is not a dev tutorial but yeah you can do it and test if you are uh interested. So yeah so we can always perfect things along the way. So here it is. So you start u by cloning the near MCP repository and you can build the uh build the project and you will be having a dis folder. So this step will

11. 4:36

    basically help you in setting up the server. So if you have a developer uh once you build it you can set this MCP server in your uh cloud config right. So you can go to cloud desktop and in cloud desktop uh you will be uh let me just see if the video is visible. Uh can you can you please confirm if my video is visible or not?

12. 5:00

    Yes it is. Okay great great because I think it was buffering or something like that. Great. So uh once you build this uh you will be having uh yes. So I think there is some issue with this video. Let me quickly see if I can fix this or not. Yeah, I think there's this.

13. 5:23

    Yeah. So me see if I can present it. Yep. Yeah. Seems like it is working. Yeah, it works. Yes, there's some bandwidth issue on my end. Not sure. Yeah, I think we can hear and see you fine. So, it looks good on my end. Okay. Okay. Great. Great. Because uh yeah, let me let me Yeah, now it's working. Perfect. Perfect. So, uh yeah, I think it's still buffering.

14. 5:55

    Yeah, it's working now. It's working. What? So, uh let me just explain uh this very quickly. So if you are uh inside clone you have a feature where you can basically add MCP server. So you can have a locally running MCP server. You can add this MCP server and yeah so here are the steps. You can basically go to settings and add this MCP server. And once you add this server you will have all the tools and all the tools will be

15. 6:24

    available inside uh your uh inside your uh cloud code client right this client. So here are some of the tools. Here are the tools that you have for chain signatures which is uh which is right now for bitcoin right so we will be seeing how you can get access to a bitcoin wallet from your cloud desktop and with just having an account so uh the best part about this is you can ask

16. 6:51

    uh uh a a chatbot context to show your uh show your wallets right so here I have a wallet that I imported and I can ask it uh anything about my wallet right so as I asked it about my near balance so it is showing me that I have some near uh and I can I can basically send a transaction so in this in this chat if you see I tried sending 0.05 05 and and you don't need to do uh any uh any like

17. 7:22

    if if if you want to do it you just tell the tell the agent and it is going to do it for you right uh in this example as you see I have imported another of my wallet which had a bitcoin uh address and last time I also used the same wallet to demo uh my uh demo my bitcoin transactions which we did with bitcoin agent so this is the same wallet and I imported it inside cloud code and you

18. 7:45

    can see that I have used this bitcoin address multiple times and by the way I don't have the key for this wallet. This is a this is a MCP uh wallet that this is MPC generated wallet right and and this is attached with my with my near account. So all of this just happens in a very chain abstracted manner, right?

19. 8:06

    And now you can have all of this through your famous through your favorite uh favorite uh LLM uh LM providers, right? So so yeah, so that's what is possible today with these uh with these tools and with this agent. You can try uh this yourself and uh we have couple of uh other chains also coming up on uh on chain abstraction uh with the near m uh near MCP server. So yeah so it's uh it's a great time this is awesome harsh um what next?

20. 8:41

    Yeah. So, uh the next thing that we have is to uh get this merged with the near main near MCT main uh branch and then build because what I'm thinking right now is that near intents has made it really seamless to trade any asset. So maybe next we can go is to add tools to directly use your bitcoin or any uh asset that you have uh in your in your near account or in in your near intents like with the with the help of near intent you can just directly swap and uh

21. 9:14

    trade uh these assets. So think uh think of it like if if you have a a AI agent based tools that can do it and bit uh is it's very much doing all of this. So we think integrating it will be a good uh step going forward. Beautiful. Well, thank you so much for joining.

22. 9:34

    Speaking of beta, uh we have the CTO here today. Um but Harsh, thank you for joining. Uh hope to see you back soon. Yeah, sure. Definitely. Thank you. Thank you. Yep. Micro, how's it going, man? You are I can't hear you. Looks like you had an audio issue. Um just make sure you enabled sound. You are unmuted.

23. 10:01

    Would you mind if we come back to you and and push to Alex until Hey, hey, hey, can you hear me now? There you are. Perfect. Yeah. Bada boom, bada bing. There you are. Yeah, I was going to say it's pretty hot today. Uh but anyways, I wanted to say um yeah, nice work, Harsh. Uh I'm not sure if we're connected yet, but if not, we should definitely connect. A lot of lot of good stuff from both the Bitcoin

24. 10:24

    agent. and I see a lot of nice stuff on MCP and that's all my up my alley. So, let's definitely connect if we're not connected yet. I will make that happen right now. But, um awesome. For those who might not know you, you're you're kind of famous in the near ecosystem. You've been around for quite a while. Um but for those who are watching for the first time, uh can you give everyone just a quick introduction of who you are, what it is, and then what you shipped this week?

25. 10:47

    Sure. Yeah. So, I'm Lu Micro. I've been around since uh 20 2020 I guess on the in the near ecosystem. Um currently the BA CTO bit is making it very easy for building chain agents or agents that interact with crypto. And uh yeah, that's pretty much what we're we're trying to do. And today I'm going to be talking a bit about our Coinbase wallet integration and what that means for developers or anyone that's interested in building uh agents.

26. 11:19

    Uh, I don't really have any slides prepared, but I have a couple topics here. I'm just going to run through a couple of them and maybe um if you're interest if you want to like stop me and ask questions, Cameron, feel free to, but essentially um yesterday we saw that Bass announced the Coinbase wallet, the new Coinbase wallet, TDA, and uh it's pretty interesting. Uh we've been playing around with it for about two

27. 11:43

    weeks. uh and we definitely saw a massive opportunity for us to try to integrate our infrastructure to the wallet or whatever uh base and Coinbase are planning to do first because the wallet is a super app. Uh that means that both your payments everything you do uh in crypto and your social the social layer is integrated into one app uh with Farcaster and all that. So that was definitely very interesting for us.

28. 12:08

    Um and the first thing we started thinking about was like all right how can we make our agents available in the wallet and how can what's the experience people will have when they they want to uh you know talk to the agents. So I think there are like mainly two sort of approaches. One is you build a mini app.

29. 12:26

    That means you can you know just build a a very normal web app and integrate that into into into the Coinbase wallet or you can make your agent available as a account like a social account whatever that you can directly send a DM to and uh and do things that way. That was the route we took. So we learned a bit about how to integrate agents. there's a couple of things you you need to to learn. So like XMTP um versus for example on Telegram you have web hooks. Uh it's quite different.

30. 12:58

    XMTP is this decentralized messaging protocol. We we didn't have any experience but uh we're still like trying to figure out a couple of things but it's working uh now. And um yeah, so agents can also have this social account which means that they don't they do not only serve as you know an assistant they can send a message to but also can posting uh can post on Farcast or can

31. 13:22

    interact with other parts of that social layer which is pretty interesting. Um so not sure if I can share my screen here but I can definitely Yeah, you can. Yep. Let me see. Okay. Yeah, I see it. Share screen. Nice. All right. So, what I'm going to show here is the developer tool. Uh the experience on the Coinbase wallet is completely different. But when you uh you know like sending a message to the agent should be as simple as creating a new direct message. In this case, I think I need to input the uh the right

32. 13:56

    address. But on the Coinbase wallet, you just need to do bita.base.eth and this will translate to uh the correct address. And then you can send a message. I'm already set up here. So, I can send a message. And you see, oh, no message was sent, I guess. Let me refresh the page. I got to say this guy. Oh, it was sent.

33. 14:25

    Let me try again. All right. So, when I send a message um in the wallet, this looks much better. But the agent reacts saying that all right I saw your message I'm processing since this is like a a DM sort of sort of style. It's not like streaming the text as it generates. Uh and yeah so how can I assist you and say what's in my portfolio.

34. 14:54

    Yeah these guys uh need to improve their developer experience. I think the message went through but we cannot really see it. In the meantime, I'm going to open beta and show you the agent I'm talking to, which is the core agent.

35. 15:12

    Yeah, I guess I'm a little curious like who's running the indexer like what exactly are they indexing across many chains? Like who's actually serving this this data uh the portfolio uh we're using Zerion for that. Uh Zerion aggregates a bunch of different chains and and everything and we're just interfacing with their API.

36. 15:31

    Got it. I Okay, it's a little odd I think the experience here, but essentially the Well, it just launched, right? Yeah. Yeah. So, uh we kind of built the proxy from uh XMTP, which is a protocol that the base wallet is using to our agents. And uh what I was going to show is like the core agent, this one is literally the same agent we're using.

37. 15:57

    And the the interesting thing here is like the wallet you connect when you're talking to the core agent from the beta website, the coinbased wallet or even XMTP developer uh tool shares the same memory. So like we're collecting these messages. We're adding that to the agent memory. So it will remember things you've asked, things you've done, your uh token preferences, stuff like that.

38. 16:19

    So that gets shared across uh all of the experiences that talk to the same agent. So yeah, so that's it. Basically, uh what were what what are our plans next? First figure out how to improve uh the XMTP layer which is uh you know right now we have a proxy. It's still I think private. Yeah, our repo is still private. But this repo will kind of be a boilerplate for anyone that wants to

39. 16:48

    make their agent available on the Coinbase wallet as well. So with I believe very simple setup. Yes, you just kind of need a couple of things. The beta agent ID, your beta API key and the wallet uh or or yeah, you know, the private the the public key or the address for your agent. So um yeah, you can just launch this thing and make it available immediately on the Coinbase wallet. I wish I was able to share my wallet, but I don't think I have a very simple way to do it here.

40. 17:20

    But yes, that's kind of it. Awesome, man. Um, so just to be clear, all the agents that are currently available in BA today can be made available in the Coinbase wallet. Yes. Yeah. Nice. We're going to be releasing documentation for this since we're still like improving a lot of the the proxy for for XMTP this week. Um, and hopefully the next week we have the documentation for for for that.

41. 17:49

    Incredible. We'll love for you to come back when that's ready. Uh just consistently, you know, highlight exactly what you guys are building. You've been shipping for quite a while. A lot of amazing stuff. So, thank you for everything and uh thanks for hopping on. Is there anything else that you want to flag before we move on to Alex from?

42. 18:06

    Not from my side. No. Thank you. Awesome, man. Thanks for joining. See you soon. See you. Bye. Alex Newman, are you ready? turn on your camera. If so, I will ping you on Telegram, Alex. All right, let's see here. Welcome. Welcome. Big problem I've got current.

43. 18:33

    Yeah, the big problem I have currently is I want to use my MacBook as opposed to my AirPod microphone. I hope you can hear me. I hear you. I hear you. Okay. Well, let's go through my Can you hear me? I hope that's okay. You're great. Good to chat. Cam, is it Cameron or Cam or what do I do?

44. 18:53

    Cam either. I don't care. I I I respond to to anything. And if you forget my name, you can call me Alex and I might even respond. So, it's um it's a joke. What I'm going to present this week hopefully super quickly is what we did this week in regards to Aurora. We did a couple of other big things, but I'm going to try to make this one minute.

45. 19:14

    I've added these videos to the stream. I don't know if that's something that we can play just to avoid the live demo risk. Yeah. All right. Sweet. See if that works. It's like you've done this before, you know, live what live demos are very Yeah. Yeah. I got the videos on my screen so I can screen share as well, but the they I don't know what I do.

46. 19:36

    Hey, Alex, real quick. Um I see two videos. I see the dashboard staking and then I see another one that is like a white screen. Which one do you want to be displayed? Job creation. Uh, let's start with dashboard staking. Okay, cool.

47. 19:51

    All right, so this is a video that they put together this week where we moved our staking over to Aurora test net. We're actually still looking at um, uh, you know, you and I have talked about that we might do a virtual network. I don't think we've announced anything about that and the specific Wait, Alex, before we hop into all this, do you mind just Yeah. Before we hop into all this, can

48. 20:14

    you just give a quick introduction of who you are, what human protocol is, because I believe that most of the audience might not know. Yeah, I know they don't live and breathe human protocol like me. How weird is that? Um, so yeah, if you go to uh let's see here, dashboard.humprotocol.org.

49. 20:31

    Um, you can see that we have done 24 billion tasks, almost 25 billion human labeling tasks for machine learning. I joke we're uh we've done more uh AI than I think anyone in crypto and um the big thing that we're we're working on now if you're people want to check out CV as in computer vision.humanprotocol.orgsock /sock. We are currently creating the world's largest animal pose database.

50. 20:59

    It's actually world's largest pose database is actually now of mice, which is kind of funny. Before humans had the record, but now our our post database about four or five times bigger. Um, and soon with human workers, we're going to release one that's 20 times bigger post database. Um, and all of the workers who uh use our system are um use our decentralized uh proof of humanity system which includes tracking which country they're in. Um, and it also helps track their reputation. And they

51. 21:28

    do tasks for us on different labeling tools and get paid in various currencies. In this case, we're mostly paying in our own governance token, which is HMT. And those workers can actually stake HMT to get better re rewards or or get into different types of pools that they otherwise wouldn't.

52. 21:46

    Also, all the infrastructure also uses the same staking interface to basically register to say like, hey, I've got new work I want to make available or I want to ask capacity to your network. Um, I think I don't know uh which example they produced, so I'm just going to press play to show. Um, why don't you ask questions about that if that makes sense? That's a good highlight.

53. 22:09

    Yeah, totally. I mean, just just to summarize, u data labeling service. I see that camera frozen. I wonder if anyone can actually even hear me. Let me I can hear you. I think it might be your signal. Oh, cool. Cool. Oh, yeah. So, sorry. So, sorry. I'm going to pause my video for a second.

54. 22:24

    See if that helps. It's probably the double video thing. Um but yeah, you can see what's going on here is this is actually uh someone registering actual oracles in as like Aurora Oracles into our system, I think, namely a job uh launcher for machine learning. Um and then we can actually just hop on to the next uh video. Let me see if I can fast forward a bit. Yeah, that's pretty much And so if you go to the hop to the next one, this is that job launching server that actually gets

55. 22:58

    registered. So now uh anyone uh will be able to detect on our dashboard this this uh job launch server it's running through Aurora launch Aurora ML jobs. And I think here they launch one of the same uh ML. Oh, this is actually a moderation job. So uh a lot of what our workers do is I would describe it as hot dog or not, but obviously it's a lot more sophisticated than that. Um, what's

56. 23:24

    cool is if you check out our current blog post, our expert workers are still cheaper than LLMs at content moderation, but our worker workers, like the people who use our system to label, are about 40 times cheaper than LLM still for content moderation, which is kind of amazing. Um, and um, the other interesting part about it is they're a lot more reliable. So, we recently put out some study about like Google Vision versus Amazon versus the other content

57. 23:56

    moderation and we actually found that Google mistook a lot of our pictures of mice in cages as pornography and so it was kind of a funny a funny little story there. Um anyway um so basically added um Aurora to dashboard for staking, added Aurora for launching jobs on a machine learning network, ran some test jobs, um starting to look at um doing a virtual network so that way we can bring this machine learning load over to Near.

58. 24:27

    And so I'm I'm going to stop there and ask any questions. Yeah, totally. Uh this is super cool stuff. Curious about you know data labeling is a huge huge huge uh industry. uh you know scale was recently pseudo acquired by Meta for billions and billions of dollars. Um you've been at this for quite a while. Uh what sort of inspired you to get into this to begin with?

59. 24:49

    Oh yeah, you know um so I helped found the world's uh largest uh capture company uh to do certain types of labeling that couldn't be done anyway like to facilitate the creation of certain machine learning devices. And um what's kind of weird about being on the inside of the industry is um you know we've been using LLMs even before open AI like a lot of the capture people were um and we see the role of aligning

60. 25:20

    humans uh and aligning models to humans uh via the internet to be kind of more important. I think a lot of people uh wonder like what are they going to be labeling and it's like well they're going to be they're going to be telling the computers like how to how to not up the whole world. And so that was really always my vision is like how do we create a universe where a world where

61. 25:45

    uh computers and machines can interact in a positive manner and I think that's a big portion of what the human protocol is is like how do we use AI to help protect workers right workers rights and how do we by you know merging machines and computers uh build new uh new new types of intelligence, new types of nations. Um and so yeah, that's what that's what gets me out of bed. Hope that make hope that's a good one.

62. 26:16

    Yeah, super cool. Um what I mean there's all these different types of data that can be labeled. Um have you been primarily focused on computer vision? Oh yeah, that's a great point. So we are actually pushing I would say anything in biology is really where our focus life science and a big part of that is just our partnership with the sock. Um obviously we do anything with computer security. So we've been working a little

63. 26:41

    bit more on um novel zero day research using LLM and people. And um we also do a lot of stuff with audio but most of our focus on audio is trying to understand the emotion behind language. We've gotten pretty good at basic translation, but even for languages like English, I don't think we've understand like the nature of uh of emotion in language. Uh and so we've been focused on that. I think partially just because I had a baby and I was just so amazed at how expressive the baby can be without

64. 27:17

    knowing any words. And it made me realize that there's this whole uh aspect uh that I think um would be good to have really cool label data sets for especially across the world and and I'm really interested to understand like how is human audio emotion different than like Bangladeshi and stuff like that or English versus Bangladesh. Wow. Sorry.

65. 27:39

    I mean and there's tons of languages out there. So u how would you even go about prioritizing that? That's very well. Well, we, you know, we are so careful about where we pick our workers from. We have kind of gone the other way and said like, well, there are all these countries that people have problems with, like the United States or China or Russia, like anyone who anyone has a problem with, we don't want those workers to start off with. Uh, which sucks because I would love to be a

66. 28:11

    worker on my system. I can't. Um and then um we um have done zero recruiting or marketing. In fact, what we have found is that whenever we need people, we have like they just pour in because there's so much uh so many people want to get paid uh a fair wage that they know they're going to get paid on time.

67. 28:36

    uh that that that part like we can get it from just basically anywhere where you know $2 an hour or less is reasonable. Yeah, I'm actually I think data labeling is probably one of the last human jobs that exists. Um but it obviously is going to independent Yeah.

68. 28:55

    Well, dependent on the types of data that gets lab. Yeah. I think how different BO is versus computer vision versus like the red teams at open AI all these are data labeling right and it's and like what's great is all those lab labelers are getting upleveled at each way at human I want to get beyond just like machine learning labeling I'm really inspired have you heard of this company public AI they seem really I'm really inspired by

69. 29:26

    a lot of the stuff that they're doing shout out to public AI and I'm I think that there's a lot of cool tasks that are still valuable for intelligence but aren't what people normally think are data labeling and I think with crypto we need to be thinking about stuff that like crypto people are interested in and so that I think is uh the future.

70. 29:50

    Super cool. Well, awesome. uh something that I would love to run eventually and I will probably reach out to um I want to run like distributed um benchmark creation uh competitions and I'm curious no not have not okay so Donsong is a friend and the most badass person ever and she uh recently assembled a bunch of people uh some people I don't know what I can say about participated but the um they uh found over a thousand novel exploits using

71. 30:26

    LLMs and um that is uh like a real gamecher I think. Um and I can tell you talking to a lot of those teams everyone was holding back their good stuff. Um so it's uh makes sense. So I think Defcon 2026 will be fully automated which is going to be wild. That's wild.

72. 30:50

    Well, Alex, super cool to have you on the show. Um, would love to have you on as, you know, these things continue to grow. Um, is there anything that you want to flag before we move on? No, thank you so much for reminding me. I'll bring more bandwidth next time. Thanks so much. Yeah, of course. See you.

73. 31:07

    All right, guys. That is all we have today. It was relatively shorter show. Um, next week we do have jam-packed agenda, so hope to see you then. Uh, I guess very high level. I did not prep this, but uh this week I've been really focused on uh a couple big partnerships.

74. 31:23

    I'll bring my my book and hopefully you'll hear more about them. Uh very interested in this like uh distributed benchmark competitions, working with a couple teams on that front. Uh working on distributed model training uh with a team uh who will also be on the show eventually. uh how do we actually uh train models inside of TEEES uh to verify content authentication and some other stuff uh focus on portable memory uh potential compute cloud uh a lot of data labeling uh ooth and login I my job

75. 31:54

    is really weird so I don't want to bore everyone with the long list of stuff but uh we'll probably give a a bigger lowdown next week um but besides that thank you all for joining uh hope to see you next week so yep shout out to your week for for facilitating as Thanks for watching. Subscribe to Nearweek.