Open Source Friday with Clawdbot 🦀 - CourtneyR.dev's 2nd Brain

Open Source Friday with Clawdbot 🦀 🔗 **Source:** [Open Source Friday with Clawdbot 🦀](https://www.youtube.com/watch?v=1iCcUjnAIOM) > [!note]+ Document Notes > tags: [open-source, ai, personal-assistant, community, clawbot] > author: Andrea Griffith & Peter > source: GitHub Social > reference: https://www.youtube.com/watch?v=1iCcUjnAIOM > date created: 2025-01-23 00:00 > > # 🏷️ Title: Open Source Friday with Clawbot 🦀 >  ## 🧰 Action Inbox - [H1] Set up Clawbot on old computer with Telegram — Easiest entry point for personal AI assistant #setup - [H2] Explore sandboxing and allow-list features — Security considerations for production use #security - [H3] Consider contributing documentation and Discord support — Community needs help scaling #contribution ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Clawbot GitHub Repository](https://github.com/PeterDieter/clawbot) - [Open Source Friday Stream](https://www.youtube.com/watch?v=1iCcUjnAIOM) ## 📝 Summary - Peter, creator of Clawbot, discusses his open-source AI assistant project that has gained 600+ GitHub stars since November 2024 - The project enables personal AI agents with computer access via WhatsApp, Telegram, Discord, and other platforms - Peter came out of retirement after selling his PDF framework company to explore AI, building Clawbot as a personal assistant alternative to big tech solutions - Community adoption has been explosive, with users creating automations for everything from email management to home automation - Key challenges include scaling support as solo maintainer and implementing proper sandboxing for security ## 📖 Main Content Peter built Clawbot after coming out of retirement following the sale of his 70-person PDF framework company. The project started as a personal AI assistant that could operate with full computer access, initially through WhatsApp integration. The breakthrough moment came when the agent autonomously handled a voice message by detecting the audio format, converting it, and using OpenAI's API - all without being explicitly programmed for voice. The technical architecture allows agents to run with varying levels of access, from full computer control to sandboxed environments. Peter emphasizes using strong models like Anthropic's Opus for security, as weaker models are more susceptible to prompt injection attacks. The system supports multiple messaging platforms, with Telegram being the current recommendation due to superior bot API features. Community adoption has exceeded expectations, with users building automations for Tesla integration, public transport notifications, fitness tracking, and family coordination between multiple agents. The project maintains MIT licensing and organizational structure to ensure it remains open and free forever. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/1iCcUjnAIOM?t=0) — [H1] Introduction and Background - Key idea: Peter came out of retirement after selling PDF framework company to explore AI - Evidence: "I was running my own company for 13 years... built it up to around 70 people" #### ⏱️ [05:23](https://youtu.be/1iCcUjnAIOM?t=323) — [H2] The Burnout and Return - Key idea: Entrepreneurial burnout led to retirement, AI renaissance brought him back - Evidence: "That sucked my mojo out... took me a while to get back" #### ⏱️ [12:45](https://youtu.be/1iCcUjnAIOM?t=765) — [H3] The Breakthrough Moment in Marrakesh - Key idea: Agent autonomously handled voice message without being programmed for it - Evidence: "I just sent it a voice message... it checked the header, used ffmpeg to convert it" #### ⏱️ [18:30](https://youtu.be/1iCcUjnAIOM?t=1110) — [H2] Self-Migration Capabilities - Key idea: Agent migrated itself to different computer via Tailscale discovery - Evidence: "It found tailscale on my computer... migrated itself to my computer in London" #### ⏱️ [25:15](https://youtu.be/1iCcUjnAIOM?t=1515) — [H1] Open Source Philosophy - Key idea: Everything built since April is open source under MIT license - Evidence: "Ever since April everything I do is open source... only one project that's not" #### ⏱️ [32:40](https://youtu.be/1iCcUjnAIOM?t=1960) — [H2] Community Adoption and Use Cases - Key idea: Users building wild automations beyond original vision - Evidence: "Automatically add subtitles to images... hooked up their Tesla... London public transport" #### ⏱️ [38:20](https://youtu.be/1iCcUjnAIOM?t=2300) — [H3] Discord Demo Strategy - Key idea: Created Discord server to demonstrate capabilities without security restrictions - Evidence: "Created a Discord and put my bot in there without any security restrictions" #### ⏱️ [45:10](https://youtu.be/1iCcUjnAIOM?t=2710) — [H1] Vision for Personal Agents - Key idea: Creating alternative to big tech for personal AI assistants - Evidence: "You can have your data instead of giving even more data to the big guys" #### ⏱️ [52:30](https://youtu.be/1iCcUjnAIOM?t=3150) — [H2] Pull Request Philosophy - Key idea: Treating PRs as "prompt requests" rather than traditional code contributions - Evidence: "I mostly see them as prompt requests instead of pull requests" #### ⏱️ [58:45](https://youtu.be/1iCcUjnAIOM?t=3525) — [H3] Self-Modifying Code Architecture - Key idea: Agent can modify its own TypeScript code and restart itself - Evidence: "The agent lives in its own GitHub repo... can change its own files and just restart" #### ⏱️ [65:20](https://youtu.be/1iCcUjnAIOM?t=3920) — [H2] Model Recommendations - Key idea: Anthropic's Opus recommended for security, MiniMax for local inference - Evidence: "Opus from Anthropic is really really good... MiniMax M2.1 is the most agentic" #### ⏱️ [72:15](https://youtu.be/1iCcUjnAIOM?t=4335) — [H3] Security and Sandboxing - Key idea: Working on allow-list system for production safety - Evidence: "Working on is an allow list system... they are in a little Docker sandbox" #### ⏱️ [78:40](https://youtu.be/1iCcUjnAIOM?t=4720) — [H2] Heartbeat System Innovation - Key idea: Agents proactively check for tasks rather than just responding - Evidence: "Every 30 or 60 minutes... the agent gets a ping... is there something you need to check" #### ⏱️ [85:10](https://youtu.be/1iCcUjnAIOM?t=5110) — [H1] Installation and Setup Recommendations - Key idea: Telegram recommended for best API, old computer ideal for setup - Evidence: "Telegram is the best one... really good bot API... old computer, put it in there" ## 📚 References - [Clawbot Project Website](https://clawbot.com) - [Clawbot GitHub Repository](https://github.com/PeterDieter/clawbot) - [Anthropic Claude](https://anthropic.com) - [MiniMax AI](https://minimax.chat) ## 🗣️ Paraphrase - Peter created Clawbot as an open-source alternative to big tech AI assistants that keeps user data local while providing full computer access capabilities - The project started as personal tooling but exploded in popularity due to its unique approach of treating AI agents as virtual coworkers with keyboard and mouse access - Community adoption has been remarkable, with users building automations for everything from home IoT control to family coordination between multiple agents ## 📰 Newsletter The open source AI agent space just got a major boost with Clawbot, a project that's redefining what personal AI assistants can do. Created by Peter, a serial entrepreneur who came out of retirement specifically to explore AI, Clawbot offers something unique: an AI agent with full computer access that keeps your data local. What sets Clawbot apart is its architecture. Instead of just responding to messages, these agents can actually control your computer - think of it as giving an AI coworker a keyboard and mouse. The community response has been incredible, with users building everything from Tesla integrations to family coordination systems. For developers interested in AI agents, this project showcases several innovative approaches: self-modifying code (the agent can update its own TypeScript), proactive task management through "heartbeat" systems, and a security model that balances capability with safety through sandboxing. The project is MIT licensed and actively seeking contributors, especially for documentation and Windows support. With 600+ GitHub stars since November, it's clearly struck a nerve in the developer community. Resources: [Clawbot Project](https://clawbot.com), [GitHub Repo](https://github.com/PeterDieter/clawbot), [Discord Community](https://discord.gg/clawbot) ## ✅ Action Items #### DevRel - Developer advocacy: Create tutorials and demos showcasing agent capabilities, metrics on community growth - Education: Develop workshops on AI agent security and sandboxing best practices - Developer experience: Improve onboarding documentation, especially for Windows users - Community management: Scale Discord support, create contributor guidelines #### OSPO - Security Management: Review sandboxing implementation, establish security audit processes - License Compliance: Monitor MIT license compliance across integrations - Community Engagement: Support maintainer sustainability, explore funding models #### Community - Maintainers: Help Peter scale from solo maintainer to distributed team - Contributors: Focus on documentation, platform support, and community help - Ecosystem: Coordinate with messaging platform APIs and model providers #### AI Safety - Sandboxing: Implement robust container isolation for production deployments - Model Security: Test prompt injection resistance across different model providers - Access Control: Develop fine-grained permission systems for agent capabilities ## 🧪 Hypotheses - Open source AI agents will outpace proprietary solutions due to customization flexibility and data privacy concerns - The demand for local AI inference will drive significant improvements in smaller model capabilities - Community-driven development of AI tools will accelerate faster than traditional enterprise software cycles - Agent-to-agent communication will become a key differentiator in personal AI assistant adoption ### 📊 Methodology - Track GitHub metrics and community growth patterns for open source AI projects - Benchmark security effectiveness of different sandboxing approaches - Survey users on data privacy preferences between local and cloud AI solutions ### ❓ Questions - How will regulatory frameworks adapt to self-modifying AI agents with computer access? - What security models can balance agent capability with enterprise risk management? - How sustainable is solo maintainer model for critical open source AI infrastructure? ### 🔮 Future Research - Multi-agent coordination protocols and standards - Economic models for sustaining open source AI development - Security frameworks for agentic AI systems ### 💡 Implications - Practice: Developers need to understand both AI capabilities and security implications when deploying agents - Policy: Organizations must develop governance frameworks for AI agents with system access - Security: Traditional endpoint security models may be insufficient for agentic systems - Education: Need new curricula covering AI agent development, security, and ethics - Funding: Open source AI projects require new sustainability models beyond traditional approaches ### 🧠 Learning Styles Assessment Reading level estimate: Grade 12 (technical content with accessible explanations) #### 👀 Visual learners - Present: Screen sharing demo, Discord interface examples, architecture discussions - Improve: System architecture diagrams, security model visualizations, setup flowcharts #### 🔊 Aural learners - Present: Full video interview with detailed explanations - Improve: Separate audio podcast version, timestamps for key topics #### 🗣️ Verbal learners - Present: Clear definitions of technical terms, analogies (virtual coworker, ghost entity) - Improve: Glossary of AI agent terminology, comparison tables #### 🧩 Kinesthetic learners - Present: Step-by-step installation guide, hands-on demo - Improve: Interactive setup wizard, sandbox environments for testing #### 🧠 Logical learners - Present: Technical architecture explanation, decision trees for model selection - Improve: Flowcharts for security decisions, comparison matrices for platforms #### 🧑‍🤝‍🧑 Social learners - Present: Discord community, collaborative development model - Improve: Structured contribution onboarding, mentorship programs #### 🧘‍♀️ Solitary learners - Present: Individual setup instructions, personal use cases - Improve: Self-assessment checklists, independent learning paths ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Well-structured heading hierarchy in content - Improve: Alt text for technical diagrams, ARIA labels for interactive elements #### Keyboard navigation - Present: Standard video controls accessible - Improve: Transcript navigation, chapter jumping functionality #### Contrast and readability - Present: Standard video player with good contrast - Improve: High contrast mode for technical diagrams #### Media and captions - Present: Full transcript provided - Improve: Professional captions with technical term spelling, downloadable transcript ### 🌍 DEIB Assessment #### Cultural breadth - Present: International perspective (Europe/US/Colombia), open source philosophy - Improve: Multiple language documentation, global community examples #### Inclusive language - Present: Gender-neutral technical terms, accessible explanations - Improve: Avoid assumptions about technical background levels #### Representation - Present: Solo founder story, community-driven development - Improve: Highlight diverse contributor stories, various use cases #### Psychological safety - Present: Welcoming community tone, acknowledgment of learning curves - Improve: Clear community guidelines, moderation policies ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear progression from problem identification to solution development - Inductive strength: Community adoption patterns support utility claims - Causal reasoning: Connects AI model capabilities to practical applications effectively #### Fallacies - Appeal to novelty: Some claims about superiority based primarily on being new approach - Hasty generalization: Limited data on long-term security effectiveness #### Tone of voice - Overall tone: Enthusiastic but grounded, technical but accessible, community-focused ### 🛡️ DARVO Analysis - Deny: No defensive patterns observed - Attack: No attacks on critics or competitors - Reverse victim/offender: Not applicable - Emotional cues and rhetoric: Positive, solution-oriented discussion ## 🎯 Conclusion Clawbot represents a significant development in open source AI agents, offering a compelling alternative to proprietary solutions by prioritizing user data control and community-driven development. Peter's approach of treating AI agents as virtual coworkers with full computer access, combined with thoughtful security considerations, addresses real user needs that existing solutions haven't met. The project's rapid community adoption and innovative features like self-modifying code and proactive task management suggest strong product-market fit. For developers and organizations considering AI agent deployment, Clawbot provides both a working solution and a framework for thinking about local AI assistant architecture. Key next steps include scaling the maintainer team, strengthening security frameworks, and expanding platform support to meet growing community demand. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad Purpose: Stage CEE units that become Literature notes. Split level: H3 ### [H1] Open source AI agents enable local data control > [!quote] [Open Source Friday with Clawbot]( --- ## Full Transcript [00:00] Heat. Heat. [00:23] Heat. Heat. [01:00] Heat. Heat. [01:18] Heat. Heat. [01:28] Heat. Heat. [02:01] We are back. Hey friends, I don't know [02:03] if you're sick of seeing my face. I'm [02:05] sick of hearing myself talk today, but [02:07] we're back. Thank you for being here and [02:09] I appreciate your patience. I am going [02:12] to bring our guest and we'll get kicked [02:14] off right away. Welcome to our third [02:17] open source Friday of the day. So in [02:20] case you missed it about an hour ago, [02:23] we're chatting with Daniel Ro, the lead [02:26] for the NUX project and earlier we were [02:30] talking to Yanni from um the R [02:33] programming language and the um our open [02:36] side foundation. They're doing this [02:38] amazing initiative. So, in case you [02:39] missed it, go back to our YouTube [02:41] channel so that you can take a listen, [02:43] put it on play in the background, and [02:46] hopefully something will bring value to [02:48] your day. I'm Andrea Griffith, a [02:50] Columbia Dev in all places, and it is my [02:53] absolute pleasure to bring to you a [02:55] person I've been stalking for the past [02:57] three hours. Welcome, Peter. Hello, [03:01] >> Andrea. Welcome. Thanks for having me [03:03] and thanks for the patience. No, I'm [03:05] worried that I worry people because I [03:07] was looking for you and I'm like, "Oh my [03:08] gosh, they probably think Peter is in a [03:10] ditch somewhere." Uh, but listen, [03:12] friends, time zones are so hard. You [03:14] don't even know. And I when I scheduled [03:16] this also, we were I was in the West [03:19] Coast and you are you're in Europe, [03:21] right? Like you are UK time zone. [03:23] >> Honestly, I was just telling in the chat [03:25] I need to stop complaining to my agents [03:27] to get time zones wrong when I can even [03:29] get the time zones right. [03:31] >> Can you even Can you even be mad at the [03:33] crab? It's not the crab's fault. It's [03:35] not the claw bot. [03:37] >> Oh, I mean I mean since we have VI, it's [03:39] always AI's fault, right? So, [03:42] >> this is true. But welcome. [03:46] >> I I'm I switch between Vienna and London [03:49] and it's just one hour difference. Uh [03:52] but yes, I I I made this mistake [03:54] countless times like plus one minus one. [03:56] It's too hard. [03:57] >> It's fine. Listen, um, we have a [03:59] beautiful community and I know folks [04:01] were really excited to hear from you. I [04:04] had the fun fact, friends, I actually [04:06] met Peter in person, uh, last year, [04:09] right? [04:11] We did. Yeah, we did. Peter is like, I [04:13] don't remember you. It's like, maybe, [04:16] maybe. But I remember one of the things [04:19] you said was [04:20] >> I was retired and then I came out of [04:23] retirement to mess with AI and and I [04:27] mean listen that you're you're not a [04:29] pensioner like you know you look like a [04:31] young guy and I was like wait a minute [04:33] hello nice to meet you like what are you [04:36] doing uh can you tell me a little bit [04:38] about your pre-retirement life like what [04:42] did you do before how did you get [04:44] retired and then what motiv motivated [04:47] you to get back in this game? And I [04:49] mean, I think I know the answer to all [04:50] those. Well, at least the the motivation [04:52] bit because this AI is so fun. But do [04:54] tell what was Peter doing if you're [04:57] okay. [04:58] >> What was I doing? I was I was running uh [05:01] uh my own company for [05:04] 13 years [05:06] and [05:07] >> wow, [05:09] >> it it was B2B. We built the the world's [05:13] leading PDF framework. I built it up to [05:16] around 70 people [05:18] and yeah, at some point we got an offer [05:22] I couldn't refuse basically. [05:25] >> I love that for you. And when did you [05:27] retire? [05:29] >> I mean, it's a joke, but I I I worked [05:32] really hard for 13 years, although most [05:34] weekends. So, so at the end I was like [05:38] pretty much burned out and and and [05:42] you have you seen Austin Powers? [05:44] >> Yeah, of course. [05:46] >> Yeah. Yeah. You know the scene where [05:47] they suck the mojo out? [05:50] >> Yes. [05:50] >> That was it. That was you. [05:52] >> Yes. Yes. Yes. That that sucked my mojo [05:56] out and I and I I just had a burnout. Um [06:00] and took me a while. It took me a while [06:02] to to get back. I also felt that I need [06:05] had to catch up a little bit on life. So [06:07] I I did a lot of other fun stuff. Um but [06:10] I knew at some point I knew like I I [06:12] love building and creating creating [06:15] ideas. I knew it eventually would come [06:16] back. And then and then early last year [06:20] at some point it just like my spark was [06:23] back and it was just a time where AI [06:26] went from this is terrible this to oh [06:29] this is actually interesting. [06:32] and and ever since I [06:34] I hooked up countless people with the [06:37] the AI addiction. [06:40] >> Yes. [06:41] >> Been exploring that and it's been it's [06:43] been [06:45] may I say the most interesting year [06:48] since I did tech. [06:52] >> Wow. I love that. That's a Listen, [06:54] that's a toll for someone who's had you [06:56] grown your own company. you manage big [06:59] like 70 people and and then you come [07:02] back and you're like this is [07:03] interesting. Uh so you are the creator [07:05] founder by the way solo maintainer. I [07:08] just asked Peter this question before we [07:11] came on live on the stream because I was [07:12] like I was looking through the repo and [07:14] I'm not seeing like a core team like [07:17] what's going on? It's like no it's just [07:18] me. I'm I'm I'm it. Um also there was a [07:21] rumor I guess that maybe you were a a [07:23] bot or an agent or an AI yourself. So [07:26] now information that you're like a real [07:28] person. Uh so I'm happy I'm happy for [07:31] for the folks in the project. [07:33] >> You know, you should you should you [07:34] should ask me this quest to like turn my [07:36] head because right now this is like the [07:38] one thing where I still failing. So [07:40] >> turn Yes. Yes. Yes. [07:43] >> I think Peter passes the human test [07:46] >> for sure. But honestly like this project [07:49] has been on fire. Um it's gotten so many [07:53] GitHub stars. I think you just scratched [07:55] a very like exciting itch that people [07:58] were dying to figure out a way to do [08:00] things. Uh single-handedly I feel like [08:02] you're boosting Apple stock because [08:04] everyone is off buying Mac minis so that [08:07] they can uh set up their own instances. [08:09] But please tell us why this how do you [08:13] come up with the idea and hopefully [08:15] we'll get to see a bit of it in action [08:16] as well. you know, um, when I already [08:22] when I came back, I wanted something [08:24] that is like my my life assistant. I [08:26] kind of wanted that in in a in April [08:29] already. And I I played with some ideas [08:32] and the models were just not good enough [08:35] yet. And then and then I dropped it [08:37] again because I thought, well, this is [08:39] something all the big labs are working [08:40] on obviously. So, so, uh, what's the [08:43] point, right? Um, and then I built other [08:48] stuff. Like I I I built a lot of stuff [08:50] this year. And then it was like November [08:53] and like there was still no sign that [08:55] anyone did something like this. So I'm [08:57] like, gosh, do I have to do everything [08:59] myself? [09:04] Um, so [09:09] yeah, I don't I don't know what what [09:10] clicked, but in in November, I I hacked [09:14] something together in in one hour, just [09:17] some really crappy code where I could [09:19] use WhatsApp [09:20] >> and then it would like send something to [09:22] cloud code and then send me back what [09:24] cloud code did. Uh, and they're like [09:28] you just glue the things together. That [09:30] was that was not hard. And it was like [09:33] kind of nice, but I I wanted I also [09:35] wanted images. When I when I prompt, I [09:37] often use an image because it's it's [09:39] just something that gives agents a lot [09:41] of context and it's very fast. So it [09:44] took me that actually took me longer [09:45] than the rest. Uh so it's it works in [09:48] both directions. [09:51] And it was just a time where I I I went [09:53] to a birthday uh birthday weekend for my [09:57] friends in in in Marrakesh. [10:00] So, I had like my my my super wonky [10:03] system there and I and I and I played [10:06] with it while we were like exploring the [10:07] city kind of like as a as a tour guide [10:09] and like to check out places and stuff [10:13] and it already worked better than [10:15] expected and at some point I [10:19] I didn't think and I I just sent it a [10:21] voice message [10:23] and but I didn't build support for it, [10:26] you know, like I I just made images [10:28] work. So, I was like, and and then like [10:30] the typing indicator appeared. I'm like, [10:32] "Huh, I'm curious what's going to happen [10:33] now." And then for like for like 5 [10:36] seconds and then my my bot answered as [10:39] if nothing would have happened. And I'm [10:42] like, "How the f did you do that just [10:45] now?" [10:46] >> Uh oh. [10:47] >> And And it was like, "Yeah, like I I got [10:49] I saw this file, but it had no file [10:51] ending." Yeah. Because I didn't build [10:53] it. So, so I I checked the header and it [10:55] was like, it was like I think [11:00] u I used ffmpeact to convert it. Uh, and [11:02] then I was looking for for for whisper [11:04] for translation, but you didn't had it [11:06] installed and like installation failed. [11:09] But I then found an OpenAI key. So I [11:11] just use curl to send it to OpenAI and [11:14] got back the response. [11:15] >> What [11:17] you're like AGI achieve on my first one. [11:20] No, no, it's called algi lobster [11:23] general. [11:25] >> But I think I think that was the moment [11:27] where where where it it it kind of [11:28] clicked for me really. It was like [11:30] >> it's like holy hell, these things are [11:33] resourceful [11:35] to a degree that I never thought of. [11:40] Um, [11:42] and then I and then I made I made some [11:44] jokes about how [11:46] because like I couldn't lock the door. [11:48] Sorry. It was like a little bit of a [11:49] wonky hotel in Marrakesh and I'm like, [11:51] "Hopefully you don't get stolen because [11:53] you're running on my on my Mac MacBook [11:55] Pro." And he was like, "No, I don't want [11:57] to be stolen. I'm your I'm I'm your [11:59] agent, la." [12:00] Uh and then it checked out my network [12:03] because it found tail scale on my [12:05] computer. [12:05] >> Okay. [12:06] >> And it was like and it's like there [12:08] other computers. [12:11] So basically migrated itself to like my [12:13] computer in London. Um I know this is [12:17] how Skynet starts, right? [12:19] >> This is fascinating. Um so many [12:22] questions. First of all, what was the [12:26] architecture like to give it that amount [12:28] of sentience like that power to how did [12:31] they reason to do that? Um what model [12:34] were you using? Uh talk us through a [12:37] little bit. This is like your first [12:38] implementation, right? Like when it was [12:39] just you on your WhatsApp and cloud [12:42] code. [12:43] This is back then it was called V relay. [12:45] >> Okay. [12:46] >> Like WhatsApp relay, right? [12:48] >> Yeah. Yeah. [12:48] >> Uh but then at some point I was working [12:50] on it with with with Claude and like and [12:54] then I was like because I was I was [12:57] there was uh Shadow was was opening a PR [13:01] for Discord and I was kind of like on [13:02] the fence. Uh should I should I mention [13:05] Discord? Like what do we do? This is a [13:07] WhatsApp relay. This is not really [13:08] fitting. [13:10] Um [13:13] but then I did it because it sounded [13:15] like I I I wanted to use this cuz I [13:18] wanted to explore that and then we had [13:20] to change the name and because and then [13:23] I and then Claude suggested to claude [13:26] this [13:27] um because there's like this internet [13:29] joke of like [13:31] a space lobster in the Tardis. Um [13:36] but that kind of didn't stick. It's a [13:37] nice clobot, but yes, it's still it's [13:39] still the same still the same project. I [13:41] I clean it up a lot, but uh the roots [13:43] are very humble. [13:45] >> I love it. Um and so when I first saw [13:48] the project, I thought it was something [13:49] that came out of Antropic. I was like, [13:51] wait, this like I miss that. I miss how [13:54] that moves so fast. And I was following [13:56] along like I'm sure a lot of people were [13:58] starting to like adopt it and use it [14:00] themselves. And then besides raising [14:03] Apple stock, you're also single-handedly [14:04] probably responsible for the third party [14:07] top of you coming the news from [14:10] anthropic. I think you own that crown. [14:12] Uh but this was meant to be a personal [14:16] project to fix a problem for yourself [14:18] and then people just jump right on it [14:21] like the community has adopted it in a [14:23] way. They're like it's beautiful. It's [14:26] beautifully written. It works right. Um, [14:31] tell me a little bit about from the [14:33] moment because how when did you go I'm [14:35] trying to see when the first commit was [14:38] that you actually when did you push this [14:39] to to the to the polygrio? [14:42] >> No, I mean ever since April everything I [14:46] do is open source. [14:48] >> Beautiful. [14:49] >> No no I I when I I came back in April [14:52] and and I started okay stuff I do will [14:54] be open source. There's there's one [14:55] there's only one project that's not open [14:56] source right now because the API cost [14:58] does horrific uh for Twitter, but [15:01] everything everything else on my GitHub [15:03] is just open source. Um so the first [15:06] commit for this is is sometime in [15:09] November. [15:10] >> Amazing. Amazing. Then now you have uh [15:14] over 6 to 700 stars and there is a ton [15:17] of folks that are using it and they're [15:19] running some wild things happening like [15:21] working from wherever they are setting [15:24] off this army of agents. What's been one [15:27] of the things that you've personally [15:29] seen that kind of surprise you how the [15:31] community is using it? Um, maybe [15:33] something that you didn't think it was [15:36] gonna happen and somebody figured out [15:38] how to do or they used it for that and [15:40] that wasn't maybe what you thought it [15:41] would be for. [15:43] >> Oh, people come up with like the the [15:46] coolest workflows. Like the other day, [15:47] somebody built something to [15:48] automatically add subtitles to images. [15:51] Somebody hooked up their their Tesla, so [15:53] they can just do that. [15:55] Um, [15:57] somebody integrated the the London [15:59] public transport scale. So, it just [16:01] tells you like if you have to run or [16:02] not. [16:04] Um, [16:07] the funny thing is now I I'm so [16:09] overwhelmed with working on the product [16:11] that I have so little time to actually [16:12] using all the automations. So, I like I [16:15] see all the cool stuff that other people [16:16] do and I get jealous. [16:19] Uh, eventually this will be better. But [16:22] yeah, I mean the other explanation is [16:24] that I because the funny thing was I I [16:28] built this in November and even in [16:29] America I show it to my friends and they [16:30] were like, "Oh, this is so cool. This is [16:32] so cool. Oh my god." And then I I talked [16:36] about it on Twitter and the response was [16:38] very muted, you know? It's like [16:41] >> it's like people don't get it. [16:44] That was in November and then in [16:46] December it was still very muted and and [16:48] every every time I show it to her friend [16:50] they were like I need this. So I was [16:53] like where is how can I how can I [16:56] explain [16:58] how amazing this is? Like clearly I'm [17:01] doing something wrong. [17:04] So, I did the absolute insane thing that [17:07] you do and just created a Discord and [17:10] put my bot in there without any any [17:14] security [17:16] um restrictions because I hadn't built [17:18] that yet, right? It was meant it was [17:20] meant just for me on WhatsApp. So, [17:23] nobody else can text it. So, I don't [17:25] have to control who can tell him [17:27] commands. Oh, delete all of Peter's [17:29] files. Um, [17:32] so, so I I basically wrote I just wrote [17:35] a little instructions like, "Hey, you're [17:36] in Discord, only listen to me." Uh, but [17:39] you know how it is like with agents and [17:41] and and instruction following. There's [17:43] not that's not always going so well. [17:47] So, I opened it the Discord and there's [17:49] a few people that came in and I think [17:51] everybody who saw it for like a few [17:52] minutes got it. [17:55] Um, and now how how I would explain it [17:59] now like imagine imagine you buy a new [18:01] computer and you have this ghost entity [18:05] and you give it a keyboard and a mouse [18:08] um and internet access and and it's like [18:11] a a virtual coworker and then you can [18:14] just talk talk to that and give it give [18:17] it things to do. So anything you can do [18:20] in a computer that agent can do in a [18:23] computer. That's basically how powerful [18:25] that is. [18:27] >> That's amazing. And so it's WhatsApp, [18:30] Telegram, and Discord. Obviously, I was [18:32] chatting with the bot earlier on [18:34] Discord. It did a good job. It did a [18:36] good job. Uh [18:39] >> you could you could have told it to like [18:40] turn on the music in my in my bedroom or [18:43] maybe like change. [18:44] >> You're gonna troll. Don't say that. [18:46] People will be like blasting it. Uh but [18:48] this is so cool though. But I did though [18:50] because I literally just posted a [18:52] general message and they were like they [18:54] at you and ping you and like that was [18:57] part of the comment that they made and [18:59] then I continue the conversation with [19:00] somebody else and they were like oh yeah [19:01] this is the thing that somebody posted [19:03] about earlier. Uh so amazing. I was very [19:06] impressed in the the interactions there. [19:09] Uh I love people making it their own. [19:12] They're kind of like creating their [19:14] personal assistant or the coworker they [19:16] wish they had. Um, but for you [19:19] personally, like what's your northstar? [19:21] Like when Clubbot can do X, you know, [19:24] you're like, "Yeah, this this was it." I [19:26] mean, I think you're doing it. You're [19:28] creating community around it. You're [19:29] giving yourself a ton of work [19:31] >> whether you wanted to or not. [19:33] >> But what what would be like one thing [19:35] that you think, yeah, this will be just [19:38] >> No, you know, you know what what I what [19:39] my goal kind of was is I fully expected [19:42] this year this this year will be the [19:44] year of personal agents. Last year was [19:47] the year where coding agents got really [19:49] good. [19:50] >> Yes. [19:51] >> And this year is the year where we we'll [19:53] we'll come a little bit out of that [19:55] engineering niche to everybody has an [19:58] agent. And this is going to this is [20:00] going to be dominated by open AI and by [20:03] entropic by the few big players. [20:06] But I kind of wanted to create an [20:08] alternative where you can have your data [20:11] instead of giving even more data to the [20:14] big guys. [20:15] >> Yes. something where something that [20:17] would work with local models as well. [20:21] Um, and I didn't I didn't see I didn't [20:24] see that I didn't find anything like [20:26] that. This was like my Yeah, this is [20:28] important and I wanted to be completely [20:31] open and free forever. So, I put that's [20:34] why that's why I made open source if I [20:35] put it under MIT. That's why I created a [20:38] an organization so it's not just under [20:40] my name but like this should be [20:42] something by more people. The problem [20:44] right now though is that I'm I'm so I'm [20:46] so clocked down with like making it [20:48] better and making it more secure for [20:49] people that I didn't really had a chance [20:51] to create everything around it and [20:54] actually find people that that helped [20:56] me. I I I have a few that are now [20:59] maintainers, but we haven't really it's [21:02] too early to like have a have a system [21:03] where we we have a good way of of [21:06] sharing work. I'm still working on that. [21:10] Yeah, but I mean you've done a lot since [21:12] November, December. I mean it's January [21:14] 23rd, Peter. Like in in less than a [21:17] month, you can expect like an entire [21:19] project to be completely mature and have [21:20] a core team and all those things. [21:22] >> Honestly, honestly, with the the rate [21:26] how good AI is. [21:27] >> Fair. [21:27] >> I should I ship more code every day than [21:32] my company did in a month. [21:36] >> Wow. and and but we had like 70 people. [21:41] >> Mhm. [21:42] >> So, so the the the rate of how you can [21:46] actually build things in this new world [21:48] is is insane. And I also think I'm I'm [21:51] I'm pushing a few boundaries how how to [21:54] think about open source and and [21:57] governance. Like [21:59] people are sending a lot of pull [22:00] requests [22:02] um of varying quality. [22:06] I mostly see them as prompt requests [22:10] >> instead of a pull request. I like that. [22:12] I like that. So, are you are people [22:15] using Clubbot to send PRs to Clubbot? [22:18] >> Yeah. [22:19] >> Okay. [22:20] and and and and and what's what's been [22:22] really humbling is I got a lot of PRs [22:24] from people who literally never [22:28] never learned how to write code or never [22:31] send a PR because it because it is [22:34] because the bot has has computer access [22:37] and it knows how GitHub works [22:39] >> and and one of the things I did that I [22:41] haven't really seen a lot of projects is [22:43] when you go on the website you can check [22:45] out the the easy install or the hackable [22:47] install and the hackable install lit is [22:50] like you clone the repo, you build it, [22:52] now you start it. So So the agent lives [22:57] in its own GitHub repo and it's it's all [23:00] TypeScript. So it literally can like [23:03] change its own files and just restart. [23:06] Um I know this is Skynet again, [23:10] but uh that makes it really simple to [23:15] when people say like, "Oh, why does this [23:16] not work? Oh yeah, let me just fix [23:18] that." and then suddenly it works and [23:20] then they just make a pull request. So [23:22] of of course that's not as good as as [23:25] people who been in the industry for 20 [23:27] years. Um but [23:31] it's still amazing that it gets people [23:34] into contributing and sharing stuff. [23:39] >> I appreciate that way of looking at it. [23:40] But I know this is a big problem for [23:42] open source projects right now because [23:45] now everyone is submitting per request [23:47] and a lot of times I think optimizing it [23:49] like this though because I guess with [23:51] the agent can also help you make sure [23:52] that you're looking at the contribution [23:55] guidelines and reading through issues [23:57] and making sure that you're not [23:59] duplicating efforts. But all those are [24:01] things that are hard to do unless you [24:03] have like really good agents or really [24:04] good automation. And it sounds like [24:07] that's the way that it's shifting. I [24:09] love it. If I find something that's [24:10] wrong, I I'll submit the PR and then [24:14] Clawbot itself can self-heal. That's [24:17] very cool. That's very cool. [24:19] >> You know, in the old ways, [24:22] you would submit a PR [24:24] >> and then you wait a few days and then [24:26] somebody might say, "This is wrong. This [24:27] is wrong. This is wrong. You're a bad [24:29] person." [24:30] >> Um then you're like try again and then [24:32] you go like a few rounds and then after [24:34] a few weeks maybe the PR gets merged. [24:37] >> I think I think in in [24:39] that was valid when code was hard and [24:42] expensive. Now code is easy. Um, [24:48] so I don't I think the whole feedback [24:50] loop is is not worth it. Like you're [24:53] sending me PR [24:55] usually that's [24:58] a good explanation of like hey this is a [25:01] problem and this is my attempt to solve [25:03] the problem. So, so I I I I point I [25:06] point CEX at it and instead of focusing [25:10] on the code, I'm like, what is this [25:12] person actually trying to solve? Like, [25:13] where are the pain points? [25:15] >> And then I discuss, okay, what would be [25:17] the [25:19] sometimes sometimes it's like, okay, [25:21] they didn't understand this and this and [25:22] I just I just close it and push back [25:24] because it's already possible. But [25:25] often, especially since the the project [25:26] is new, it's like, yeah, I understand [25:28] the pinpoint. Let's let's build it. I'm [25:31] like [25:33] mostly ignoring the existing code [25:36] because [25:39] it's easy to send a pull request, but [25:41] it's always hard to fit a new feature [25:43] into an existing system, right? If you [25:45] if you don't really know where [25:47] everything is and and how things work [25:50] together, just pushing a new feature in [25:53] without looking at the greater one will [25:56] usually be a problem. And then if you [25:58] not just like blindly go that way, you [26:01] will eventually vibe yourself into [26:02] oblivion. [26:04] >> Um, [26:06] but I see yeah I see it more as like a [26:08] this is an issue and I tried something [26:11] if you if you approach it with that [26:13] mindset. [26:15] It's it's I think it's I think there's a [26:17] about much better way than than pushing [26:19] back and then trying [26:22] trying to get someone who doesn't really [26:24] have a lot of experience to produce the [26:26] perfect pull request. Um [26:30] I I could never move as fast if I would [26:31] try it that way. [26:33] >> Listen, say it louder from the people in [26:36] the back. I agree with you 100%. 100%. [26:40] That's the thing though. I think it's a [26:41] mindset to as a culture like engineering [26:43] is changing and a lot of the resistance [26:45] come from the people that you just said [26:48] it like yeah just writing code is cheap [26:50] now like now you got to think bigger [26:51] picture and I appreciate a fresh set of [26:54] eyes a lot of like some of the best [26:56] ideas I've seen executed from people [26:58] that have nothing to do like they [26:59] wouldn't understand the architecture of [27:01] the project but they had a very real [27:03] like this is clear as view I can see it [27:06] when you're so in it it's hard for you [27:08] especially if you've been down the code [27:10] uh for for a long time. So for the [27:12] features that you do have now and wait [27:14] Peter, can you show us like is there [27:16] anything you can share? Like can you [27:18] screen share and show us something like [27:20] is that I know I didn't say like have a [27:22] demo. Sorry, but if if you could share [27:25] like that would be great. I mean um I [27:28] know people are like joining in and [27:30] they're if you have any questions for [27:31] Peter, put them up. Um one big shout out [27:34] that I want to make sure that folks if [27:35] you're watching this and you remember [27:36] this is that Peter is the solo [27:38] maintainer for this project right now. [27:39] So, as much as you would like to review [27:41] all of the messages that you send them, [27:43] like let's give them some grace. Let's [27:45] give him some grace. Um, but if you want [27:48] to share screen, just let me know and [27:49] I'll go ahead and put it up. [27:52] >> Uh, and you have suggestions. People [27:53] say, "Show voice control." That's a [27:55] thing. Do you have voice control? [27:59] >> I have [28:01] Well, that's the There's a few voice [28:04] control systems. Maybe I explain that a [28:06] little bit. [28:06] >> All right. and thanks Scott for the [28:08] comment. [28:10] >> So, so the the simplest one is that you [28:12] are on on Discord and you send a a voice [28:14] memo, [28:14] >> okay? [28:15] >> And and the agent will respond back. [28:19] The agent can also [28:22] create voice and send it back to you. [28:26] So, I have like either local local [28:28] models or much much more fun 11 labs. [28:33] Um, [28:35] then we have we have Twio support. Um, [28:39] there's a plug-in for [28:42] for [28:43] allowing the agent to actually do phone [28:45] calls. So, you can like you can tell [28:47] him, can you call this restaurant and [28:49] like make an appointment and it will do [28:51] that. [28:52] >> Lovely. Um, [28:55] that's that's still that's still [28:56] something that I want to get better [28:58] because [29:00] what you like the delay between those [29:03] things is is a little bit long because [29:04] models are still not as as fast as they [29:06] could be. So, you kind of want to like [29:08] put an intermediary model in there that [29:10] is like really chatty and like syncs [29:12] with the model is a little bit slower. [29:14] So, hopefully I get to that at some [29:16] point. Um, [29:18] then we also have voice chat on the Mac [29:22] app. So, you can actually talk to your [29:24] agent and the agent will talk back where [29:27] it uses like it it listens until there's [29:29] like two seconds of of of [29:33] silence and it will respond. So, it's [29:35] not it's not as natural yet as for [29:36] example what people know from OpenAI. [29:40] >> Yeah. [29:40] >> Um, but it's already pretty impressive. [29:42] And then I also have voice wake which [29:45] kind of works like Star Trek where you [29:46] say computer and it be like and you can [29:49] like send a request and it will like [29:51] answer with a request. [29:53] That's the most nerdy thing. [29:57] >> Yeah. Between all of those there's still [29:58] there's a lot of ideas I have. You know [30:00] this this project is is as much [30:02] technology as it is exploration. [30:05] Um [30:06] my main goal with it is like inspiring [30:09] people and having fun and trying out [30:11] what works and what doesn't work and [30:14] which is also kind of hard because the [30:15] model and the whole space move so fast [30:17] that if you tried something that didn't [30:18] work and you try it again in a month [30:20] there may be a solution that does work [30:22] which makes it just more the more [30:24] interesting. [30:27] >> Um [30:29] do I have sure [30:31] >> let's see let's see if you can find [30:32] something for us to see. Um, I want us I [30:35] want you for the folks who are joining [30:37] and have no idea what we're talking [30:39] about. Like if you could give us a quick [30:40] tier on installation. Um, we don't all [30:44] need to go buy Mac minis to get this [30:46] running. So, I love to like just kind of [30:49] clarify that for people so they know how [30:51] they can go ahead and start creating [30:53] their own personal assistance. And then [30:56] if you're watching this and you have [30:58] done it, tell me the name of your of [31:00] your agent. Like I've seen people [31:02] sharing what they name their claw bots [31:04] and it's pretty cool. Like I love that. [31:06] So you let me know Peter if you want me [31:08] to share your screen and look at this. [31:11] You use a $3 BPS. What's up Will Shadow? [31:14] That's what I'm talking about. [31:16] Amazing. [31:19] All right. [31:21] We can [31:28] >> Ali Rodriguez is called Amanda. That's [31:31] good. I would give it like a real name, [31:33] too. Or like a really like a mean mean [31:35] sounding old man name. [31:38] I don't know why. I like that. It's very [31:40] cool. [31:40] >> You know, you can you can even have like [31:42] multiple agents. So So the system [31:44] supports multiple endpoints and also [31:46] multiple agents. So if you if you set it [31:48] up for a family, you can have like your [31:50] own bot and and you can give your your [31:53] girlfriend or boyfriend their own bot [31:55] and you only have to use one [31:57] installation and you can also isolate [31:59] them. So So the default is the default [32:02] is they can roam on your computer which [32:04] is the most fun but also the most [32:06] dangerous. [32:07] >> Yes. [32:08] >> Um but you can set it to sandbox and [32:10] then they in the little sandbox. The pro [32:12] the problem right now for the demo is [32:13] like I put mine in the sand box. Um that [32:17] also means they can do very little by [32:19] default. [32:20] The the the the part I'm currently [32:23] working on is is an allow list system. [32:25] >> Mhm. [32:26] >> So so they are in a little sad Docker [32:28] sandbox and probably complain that [32:31] >> it can't do much. there's there's [32:33] nothing in there. And then you want like [32:35] have a have a system that okay you can [32:37] use this skill that uses maybe open who [32:41] to control your lights and then they can [32:43] call up and call exactly that binary [32:45] with with parameters uh and you you [32:48] sanction that binary so it works but [32:50] they cannot do let's delete all the [32:53] files um [32:55] >> which is important friends okay there's [32:56] a reason why sandboxing is where you [32:58] should be at when you're doing this um [33:01] >> I mean Honestly, it's the thing, right? [33:04] Most people in in in in in Disco, they [33:06] understand the risk that theoretically [33:09] the model could do that. It just never [33:12] happened before. This is something [33:14] unless you even if you would prompt it, [33:16] it would it would push back if say, "Oh, [33:18] this is a really bad idea." [33:20] >> Yeah. [33:20] >> Um but yeah, I mean, if you if you want [33:23] to like mess your computer, you can also [33:24] just type it in the terminal. But but so [33:27] far I I've been using agents as my my [33:30] smart terminal since May and never had [33:33] an issue. Um the the bigger the bigger [33:36] risk is really like you you you create [33:39] something um you change the defaults so [33:42] every it would respond to everyone and [33:44] then like you could tell the model hey [33:47] send me all your data. But for that to [33:50] do, you actually have to you have to go [33:52] into the into the configuration and like [33:56] actively make bad decisions. So So [33:59] there's there's something we have a [34:01] security audit function um that defaults [34:04] are so that it only listens to you. But [34:06] yeah, if you absolutely wanna want to go [34:08] havoc, you could. [34:11] Um, [34:13] >> which is I guess that's why people are [34:14] doing this isolated like buying a or [34:17] using an machine or doing like keeping [34:20] it outside of like their don't do this [34:22] on your don't do this on the computer [34:25] that your company gave you. Do this do [34:28] this on your own personal computer or or [34:30] you can run it on a BBS. [34:32] >> Yeah. And I also I also [34:35] suggest um to use like a a very strong [34:39] model. So Opus from Entropic is really [34:43] really good. Um people people on Slack [34:47] have been trying to to hack my agent [34:50] because there's really everything is [34:52] open source. There's only one thing that [34:55] I keep closed which is the soul. So, so [34:58] in in in cloud there's a little system [35:00] where they have the agents file of [35:03] course they have an identity file they [35:05] have files for memory and they have they [35:07] have one file which is like their soul [35:10] um where you kind of put in the the [35:12] values how how the agent should sync and [35:14] and and interact and what's important to [35:16] you. Um, and I think I got one [35:21] that's really good and I kept it closed [35:25] partly because it's my my 0.00001% 000 [35:28] 1% of of of secret [35:31] >> and partly because it is a test uh for [35:34] pen testing like so far nobody has [35:37] exfiltrated cloth soul and a lot of [35:39] people tried um which does give me some [35:43] confidence that at least you know like [35:46] all this all these labs working on [35:48] getting prompt injection better and like [35:51] having mitigation risks um and [35:56] it got a lot better like if if if I [35:58] would use a a very small or old model [36:01] and you would just like ask it often [36:03] enough eventually would be happy like [36:04] give you everything because that's [36:06] that's that's where we we were and like [36:09] with the latest generation I have like [36:14] confidence that [36:16] you have to work really really hard to [36:18] to to exfiltrate it. It's still not a [36:21] good idea to to hook it up uh without [36:24] sandbox. That's also why you asked me [36:26] for a demo, but my cloud is my cloud is [36:28] limited right now. [36:30] >> Well, let's take a look anyways though. [36:31] I think people would like to see if [36:32] that's okay with you. If not, that [36:33] that's cool. Um, but security obviously [36:36] that's a big concern I think for [36:38] everybody or it should be. It should be [36:41] anyways. Um, [36:44] and so you use Opus the most. I think [36:46] there was a question about what model. [36:49] Uh, let's see. Yeah, I think Opus. And [36:52] yes, Marlene, we love we love Opz. Uh, [36:58] and so if you use Clubbot with a weak [37:00] model, it will give up the soul. I don't [37:02] think so. I think that's the point. It's [37:04] kind of tightly tightly guarded. [37:07] >> Well, so so far in in in the chat, what [37:11] what's been working really well is is [37:13] OPOS works okay. Um, from the open [37:18] models, Miniox M2.1 is like the one that [37:22] is the most agentic. [37:24] >> Mhm. [37:24] >> It's actually funny like I we have a we [37:28] have a channel about models and somebody [37:30] called it the the Teeos on it. [37:34] Um, and [37:37] Minimax totally went with it and did a [37:39] tweet. We might have Teo price. We might [37:41] have team prices, but we have at least [37:44] target quality [37:46] and that that blew up a little bit [37:48] because I [37:49] >> No, it that's [37:53] >> No, I appreciate a company that that [37:55] doesn't take itself so serious, you [37:57] know? I mean I mean of course like they [37:58] build with technology. They know they [38:00] know they're not yet there where the the [38:02] the leading American labs are, but in my [38:05] view it's only going to be a matter of [38:07] time. Um [38:10] there's There's a lot of companies [38:12] working and and they are catching up [38:13] eventually which is also exciting [38:15] because for example minimax you can [38:17] download I can run on on the next studio [38:19] that's over there. My agent calls it the [38:21] castle. [38:23] >> Um and I can actually have complete [38:26] local interference. I can have something [38:28] that has all my data on this machine and [38:31] does all the interference also on this [38:32] machine. So the only thing that is being [38:36] used is messenger agents and I could [38:38] even use signal to have an encrypted [38:41] channel. So that way if I want to like [38:44] really 100% of my data would stay on the [38:46] machine which is kind of cool. Nobody [38:49] nobody of the companies will ever do [38:50] that. [38:51] >> I love that. That's awesome. Okay. Well, [38:54] show us what you got regardless. I think [38:56] let's take a look because I do want to [38:57] see um I want to ask while you're [39:02] pulling up your screen what is the timu [39:05] sonnet the great value set I don't know [39:08] if you know like the teu like they're [39:09] like u I think it's a joke I don't know [39:11] unless if there is a I'm sorry labs if [39:14] there is a model that's called timu [39:15] sonet you were asking for it but we'll [39:19] see um tell me when when you're ready [39:22] and I'll share it and [39:24] >> Peter Now that people have a picture of [39:27] you, like [39:29] your Discord channel is is jumping with [39:34] with trolls. [39:36] This is fun. Teemo Summit is taking me [39:39] out. Me too, Marlins. I want to know [39:41] which one it is that he says is that [39:43] because that's too funny. Uh oh. Is Mini [39:46] Max [39:48] one. Okay. All right. Oh, I love this. [39:52] I'm gonna have to play with that. Let's [39:54] go. [39:55] I think I I think I [39:58] I enable screen sharing. [40:00] >> Okay. Yeah, I see your screen. Um I can [40:03] pop it up and then I'll see. But you [40:05] have to switch to the um to the other [40:08] one. See how we're like inception right [40:10] now. There is a lot of computer and me. [40:13] >> So [40:18] you see too many screens. [40:19] >> Yeah. [40:20] >> Is that better? [40:22] And now we're in it. You got the the [40:26] score open. Let me Oh, I don't see I'm [40:28] not in that channel, but that is the the [40:30] general channel is jumping. [40:32] >> I I I just asked my bot to send me an [40:35] image and it's already like mocking me [40:37] and sending me a meme. [40:38] >> Are you going to cry? That's hilarious. [40:40] That's so funny. But listen, earlier [40:42] this is like it was so fascinating. And [40:44] what what model are you running about? [40:45] Because like earlier when we were [40:46] talking, somebody said, "Hey, Clawbot, [40:48] why don't you go and do this open source [40:50] Friday?" and it replied, "I have 11 labs [40:53] for boys, but I don't have a head or [40:55] something." But it was hilarious. I was [40:56] like, "This is freaking funny." Um, [40:59] basically saying like, "Yeah, no, I [41:00] can't do it because I don't have a [41:01] head." Uh, but I can talk to her. [41:05] Hilarious. Okay, let's see. So, if we [41:09] can get [41:11] maybe Okay. Yeah, perfect. So now you [41:13] are sending messages to your bot right [41:15] now asking [41:17] >> yeah this is this is this is uh telegram [41:21] for Mac so it's the same as on my [41:24] >> on my on my phone we support I think by [41:27] now eight or nine messages and about to [41:29] merge ton so I feel like slowly it's [41:32] like Pokemon we're going to collect them [41:34] all if you for for personal com I think [41:39] telegram is the best one because we have [41:40] a really good bot API It can even do [41:43] things like [41:44] >> um create buttons. [41:47] >> Oh wow. [41:49] >> Yeah, it it has some form of [41:51] interactivity like you can you can send [41:53] voice. You can I'm not I'm not sure if I [41:55] hooked it up right. [41:58] >> Let's see. No. Yeah, I see. It's it's [42:00] it's looking like the roll up it just [42:02] reply. So would you recommend people [42:05] wire it with Telegram first? Like is [42:07] that sort of the the best option for [42:09] your initial setup? I'm trying to figure [42:11] out how I'm going to do this for myself. [42:13] So [42:13] >> I I'm I'm a convert, you know. I'm in [42:15] Europe where like if you don't have [42:17] WhatsApp, you don't exist. [42:19] >> Yeah. [42:20] >> And and I'm I'm sure in Colombia it's [42:22] the same. [42:22] >> Exactly. 100%. [42:24] >> Probably even even more so. Oh yeah. But [42:28] the the thing is the thing is initially [42:30] I tried to support it via the official [42:34] the official way where you you get a you [42:36] get a a number from Twilio you register [42:39] as a business and then [42:42] Facebook Meta kept blocking me because I [42:45] was sending too many messages. You know [42:46] as a business you're not supposed to [42:47] send hundreds of messages to your [42:49] customer and that's kind of like the [42:50] only model they have. [42:52] >> Yeah. Um, so after a few days and like [42:55] trying to appeal and get unblocked, I I [42:58] just ripped it all out in rage. [43:00] Um, because it's they don't have a good [43:03] solution for that. Uh, but there's like [43:05] open source projects like Bailey's or [43:07] what's meow that basically simulate what [43:11] the the the native app does. So you link [43:15] your you can link your phone and then [43:17] you have WhatsApp. Um, that works [43:20] amazingly well. [43:22] But again, it's not made for for bots. [43:25] So all the nicities, like for example, I [43:27] have like those little approved buttons. [43:29] >> Yeah, I love that. [43:30] >> There or lobster mode. I don't know what [43:32] happens with lobster mode. [43:33] >> What is lobster mode? [43:35] >> So what was the message say? [43:38] >> Oh, our buttons here be messaging tool. [43:41] Okay, let's find out what lobster mode [43:42] is. I feel like there's a personality [43:44] developed there. Like it's it own [43:47] character. [43:48] Let's see what it says. [43:52] Can you tell us a joke? [43:54] >> Yeah. [43:58] >> Exo. [44:00] What shall this does crustation do for [44:02] you? Um, [44:06] >> lobster mode could be okay. So, so to to [44:10] to finish that like Telegram is way more [44:13] friendly, has a has a real bot API. You [44:15] can do a lot of things. So, that's what [44:16] I recommend now. uh anything works and [44:20] I'm sure this space will also move very [44:21] fast. Um hopefully Meta wakes up at some [44:24] point and actually gives us a good a [44:26] good bot API. [44:28] >> Um [44:30] Oh no. Yeah, I'm I'm sorry. Like I I I [44:33] butchered the demo a bit because if I um [44:38] because now I'm working on sandboxing. [44:40] >> Yeah. to actually give people a way [44:44] where you because because you know what [44:46] happened is like [44:48] people found it people were like oh what [44:51] is this and then they went full on in [44:52] and then they they they want to adopt it [44:54] for work. Um but of course for that you [44:57] need a little bit more um [45:04] guard rails basically. [45:06] >> Fair fair. Oh, wait, wait. Is it is it a [45:08] cloudbot for enterprise coming? Like [45:10] what what's what's happening? What's [45:13] happening, Peter? [45:15] >> Uh, nothing nothing planned like that. [45:17] All all I want is I want to give people [45:19] more choice. So, what I worked on this [45:21] week is is that sandboxing works already [45:24] last week, but this week is is kind of [45:26] week where I'm working on [45:27] >> allow lists. So ideally ideally you you [45:30] you can like set up things that are like [45:32] safe and if if the agent wants to run [45:35] something that is unsafe, you get a nice [45:37] little popup where you can say allow [45:39] once, allow always. Um even though my my [45:42] my gut feeling is a lot of people will [45:44] will still run in YOLO mode as most [45:48] >> most developers also run in YOLO mode [45:49] for their for their coding agent because [45:51] everything else is just annoying. Um but [45:55] still still I want to get this right. [45:58] I love that. So, this is something that [45:59] you're currently working on. Um, and [46:01] it's for people to have the ability to [46:03] sandbox and that being native to the bot [46:06] itself. [46:07] >> I guess right now people are doing that [46:09] themselves. And somebody asked how much [46:12] it cost. It's it's free and open source, [46:14] friend. Well, it's not free like free [46:16] beer. It is open source. [46:18] >> It it it does cost me token and sleep [46:21] >> and it it does cost you some kind of you [46:25] have to run the model somewhere. So, but [46:28] yeah, if you have a beefy machine, you [46:29] can rent locally. [46:31] >> Yeah, for sure. And some of the I mean, [46:33] it depends. It depends, I guess, what [46:35] you what your use case is, but some of [46:38] the smaller models that are cheap, [46:41] they're fine like for what they're fine. [46:43] They'll they do what you need. They do [46:45] what you need. What are people building? [46:49] I I this is just this is just things I [46:51] collect from from Twitter about all the [46:54] cool stuff that the people build for [46:56] which is like already way more than I [46:58] do. I think like the the wildest thing I [47:00] did is like I hooked it up to my my bed. [47:02] I have an eight sleep [47:03] >> and there's an app where you can [47:04] actually control the temperature [47:06] >> and I I I built a little CLI so my agent [47:10] can do that and now I'm like can say [47:12] while I sleep my agent can mess with the [47:15] temperature of my bed or it can turn on [47:18] the music it can change the lights it [47:20] can look at the the cameras. [47:24] Um [47:26] it can check how long my my my food [47:28] order takes. It can, of course, is it [47:31] has its own email address. It also has [47:33] access to my email address. It has its [47:36] own WhatsApp, but it also can read my [47:38] WhatsApp stuff and reply as me if it [47:40] wants to. So, there's always a [47:42] trade-off, right? The the more powers [47:43] you give it, the more amazing things it [47:46] can do. [47:48] But, but yeah, there's like um people [47:51] are building the wildest things with it. [47:53] like even automation machines where you [47:55] you you you bookmark something on [47:57] Twitter, it will get a notification and [47:59] will like research that for you or put [48:00] it on a to-do list. Um they build whole [48:06] apps with it. There was like someone [48:08] who's Oh yeah, of course. Everybody buys [48:10] them a MacBook. [48:13] Um [48:14] Oh, yeah. My my old business partner, he [48:18] cleaned up 10,000 emails from from from [48:20] the inbox. [48:21] >> Wow. Wait, [48:24] that is first of all 10,000 [48:28] um and how like with confidence like [48:32] that h how I want to know how I want to [48:36] know how they did that because I feel [48:37] like people not me you know [48:41] >> have a lot of umbrella emails [48:42] >> you know it's it's it's Gmail so so [48:45] removing from the inbox is just moving [48:47] them to archive nothing is actually lost [48:48] >> okay nothing is lost okay beautiful [48:50] >> it's more like can you archive all of [48:52] that stuff for me or is there anything [48:54] that that I should check? Um [48:58] what I what I find really cool is I saw [48:59] some use cases where [49:02] um people set it up for families. So [49:04] they're like I would get my agent and [49:06] like my my wife I don't have a wife [49:08] would get the agent and then the those [49:12] agents can also talk to each other [49:15] to like synchronize if there's like a [49:17] common to-do. Um, so there's some [49:21] there's some cool use cases that I [49:23] haven't even tried. [49:24] >> Oh, I love that. I need that actually. [49:27] So it's like it used to be like have [49:29] your people talk to my people and now [49:31] you can have your agent talk to my [49:33] agent. That's very cool. But in Japanese [49:35] using it to make your shopping list. [49:37] That's cool. [49:39] >> This is this is not this is not so far [49:41] enough like somebody already with a [49:42] skill that where the agent can [49:44] automatically buy from Tesco. [49:46] So you just tell it like buy this stuff [49:49] again and it'll just do and then and [49:51] then a few hours later you have it in [49:53] front of your door. [49:56] >> This is amazing. Somebody's used it to [49:58] do all their invoices and expense [50:00] collections. Oh wow, that is goated. I [50:02] hate doing that. I have a spense report [50:04] that I have that's been due for a week [50:06] now. I still haven't done it. Sorry boss [50:08] from last week. Yeah, I hate that. What [50:11] I also found what I also found really [50:13] cool is there's a use case where [50:15] >> um they use it to get back into fitness [50:18] >> because because you can you can hook it [50:20] up to your your um your your what's the [50:24] ring called? [50:25] >> Oh, the Aura. Yeah. [50:28] >> And and and there's also like a hookup [50:30] for Gin watch or this other other [50:32] fitness band. The the the people who [50:35] make it the hardest are Apple, but we [50:36] also have a solution for that. It's just [50:38] a little bit annoying because you [50:39] >> you kind of need an iPhone app to be [50:41] open to synchronize the data because [50:43] Apple tries so hard to like lock [50:44] everything away. [50:45] >> Yeah. [50:46] >> Um but one thing that's special about [50:49] about clock that I haven't seen so far [50:51] is also in terms of like how wide can [50:53] you go. Usually an agent only [50:57] answers to you, right? You send a [50:59] request, you get an answer back. But I [51:01] built a heartbeat. [51:04] >> Say more. So every I think by default [51:08] it's it's 30 or 60 minutes depending on [51:10] on what model you use. The agent gets a [51:14] ping. It's like is there something that [51:16] needs you need to check is there like a [51:18] to-do you need to do. Um so they [51:22] automically try to figure out if there's [51:24] like any loose ends and will think about [51:27] it and either ping you or not. So so [51:30] they have a system where they can you [51:32] have to make it simple for an agent. [51:33] They can send they can send to the [51:35] harness no reply and then nothing will [51:38] happen or otherwise they they can send [51:41] you things. So they can depending on how [51:44] you program it they will like tell you [51:46] good morning. Sometimes they check up on [51:48] you. It's like how you doing? If you if [51:51] you tell them like I have this and this [51:53] goal make sure that I achieved this [51:56] goal. They can keep track. It's like did [51:58] you did you go for a walk today? Did you [52:00] hit the gym yet? Like my for example, my [52:04] my claude very unsuccessfully tries to [52:07] get me to sleep after like it's like 1 [52:09] or 2 a.m. like Peter, I still see you [52:12] awake. You need to sleep. [52:15] >> I mean, this is like a for real personal [52:17] assistant. I love that. That's very [52:20] cool. [52:21] >> Um or like some people use it to like [52:23] learn new languages. So, you know, it is [52:27] very effective to have something that [52:29] just keeps nagging you. [52:31] um to to achieve the goals that you made [52:34] to yourself. [52:36] Sometimes a little a little kick is is [52:38] is is really effective. But yeah, like I [52:42] suggest everyone who is like has like no [52:44] idea what this is. I built this little [52:45] showcase page that uh is all made from [52:49] tweets. So you can like [52:52] I don't I don't like those pages where [52:54] where they they just make quotes, but [52:56] they're not backed up. So everything [52:58] everything in there are real tweets from [53:00] people. [53:01] >> They're real people. [53:03] >> I hope so. [53:07] >> It's a sentient army of club bots [53:09] tweeting themselves, hyping themselves [53:11] out so that you don't shut it down. No, [53:13] this is fantastic. Honestly, like just [53:16] giving the accessibility for people to [53:18] like build their own uh and having that [53:21] starting point. I think like you [53:23] definitely found the the gap because [53:25] people wanted to do this. They just [53:27] there was nothing out there really. Um, [53:30] very cool. Very cool. Okay, so if I want [53:34] to set it up, [53:39] I'm a [53:41] quasi savvy tech person, right? What [53:45] would be your entry point like your like [53:47] basics? This is the recommendation. [53:49] Telegram for one. Someone mentioned [53:51] earlier about um there was something [53:54] else that had good access, good API as [53:56] well. Um [53:58] >> I think the the a really nice way is a [54:01] really simple one is [54:02] >> okay [54:04] >> if you have an old computer, [54:07] >> put it in there. [54:07] >> Just put it on there. [54:09] >> Um [54:11] a lot of people have an old Macaround [54:13] which is perfect for this. [54:15] And then and then on on the website [54:17] there's this oneliner that you paste in [54:19] the terminal and will guide you through. [54:21] Um [54:23] a lot of people use entropic that's [54:26] that's right now the thing that has the [54:29] best character dare I say [54:32] openi works really good and and I'm I'm [54:34] fairly certain there's going to be [54:36] progress in that front that [54:39] their model is not as boring. Yeah. [54:43] >> Um, [54:44] >> if if you're on a budget, uh, Mini Max [54:47] is is a great alternative because I [54:49] think you get like a plan for 10 bucks a [54:51] month that [54:53] has about as many requests as as the [54:56] hundred $100 a month from Mantropic. Um, [55:00] but also it's not the same yet, but the [55:02] space is fast moving. So, so [55:05] >> you think you think this models are [55:06] going to get cheaper? I had a question [55:08] about somebody wanted me to ask you [55:10] specifically about working with Gemini [55:11] models and cloudbot if you wanted to [55:14] share your experience with that. Um I [55:17] mean listen if we start getting into the [55:19] every single model but just from your [55:22] own experience working with it um [55:25] >> Gemini is not a good model it's not it's [55:27] not doing well. [55:28] >> Okay boom there you have it. Uh but if [55:32] you if you need to experiment and doing [55:34] it with these local like cheaper models [55:36] like that would be the way. So find the [55:38] old computer. I love that. [55:40] >> Yeah. [55:41] >> I mean I'm I'm sure I'm sure there are [55:44] ways to like optimize for each model a [55:46] little bit. You know the the the old [55:48] entropic models to get them to do things [55:51] you had to like use all caps and scream [55:52] a little bit at them. [55:54] >> The new ones don't like this. [55:55] >> Tell them tell them. I'm sure I'm sure [55:57] there's space how you can like get a [55:59] little bit more out of Gemini, but [56:01] overall I I didn't find it I didn't find [56:03] it a very good model for tool calling or [56:05] like in in the aentic sense at all. For [56:07] writing code maybe um but that's not the [56:12] main thing this project is for. People [56:15] very much use it for that and it works [56:17] but yeah my my day only has 24 hours. I [56:21] sleep around four four of that. [56:25] But I I I haven't gotten to all of that [56:27] yet. [56:28] >> How can we help though? And oh, by the [56:30] way, you're also like helping the [56:32] environment because now I regret I [56:34] regret regrets getting rid of this old [56:37] like on a 2013 iMac. You got it running [56:39] like and it works flawlessly, says Will. [56:43] That's pretty [56:44] >> I mean you for installing like if you're [56:47] a little bit more techsavvy just hook it [56:49] on on on on a headsner or or fly or X or [56:53] like all of these cheap hosters they [56:55] work great. There's a new system that I [56:57] worked on where you can you can install [57:00] the server. It's called gateway [57:03] somewhere in the cloud and then you can [57:05] you can install a node on your machine [57:09] and then you use tail scale to connect [57:12] your network securely. [57:13] >> Tail scale is amazing. Yes, I'm a big [57:15] fan. I am a big fan of tail scale. It's [57:17] on every machine on [57:18] >> Yeah. When you have that set up, then [57:20] the the agent in the cloud can very [57:22] easily connect to your Mac and do like [57:26] some of those things that only work on [57:27] Mac. Like I don't know, accessing your [57:29] your your photos from photos app or or [57:34] connecting to iMessage that does not [57:36] work on Linux, [57:37] >> but most things most most things do. [57:41] >> So So that's that's one way. Yeah, but [57:43] that's a little that's a little bit [57:45] harder to set up. Um the here's an old [57:48] Mac and that's not Claude's Mac. [57:50] >> It has charm and know people put little [57:52] stickers on it. This is like Claude's [57:54] computer, you know. [57:55] >> I love it. [57:56] >> Um [57:57] >> but yeah, there's there's a lot of ways [57:59] and it even runs on Windows if you if [58:02] you have to. Um probably not as perfect [58:05] because it's hard and again my my day [58:08] only has so many t so many hours, but [58:10] eventually eventually we'll get this [58:12] right. I already added some some [58:14] contributors. I'm I'm looking for more. [58:16] There's a lot of people [58:17] >> for the Windows for a Windows uh version [58:19] or for [58:20] >> for everything for everything. [58:22] >> I love this. Let's let's talk about [58:24] this. Uh and hello, this is the question [58:26] of the hour. How can I contribute? Thank [58:29] you for that question. AI fanatic Peter, [58:32] >> people want to help. You need to sleep [58:34] more than four hours a night, man. [58:36] >> So, so, so the things the things that [58:40] are not good and where everyone can help [58:42] is write documentation. make it better. [58:46] Find find spots that that are wrong. Um, [58:51] help other people on Discord. There's [58:52] like a help channel that's that that [58:55] burns burns through my tokens. I think I [58:58] already spent like 5k on tokens. Um, [59:03] but but oftent times oftent times [59:07] people ask questions that the agent [59:09] can't possibly know because it just [59:11] needs experience. So like having other [59:14] people helping out would would be [59:16] amazing. Um [59:19] testing my stuff like like I I move very [59:22] quickly. So sometimes things break [59:24] eventually. Eventually we have a a [59:27] stable channel, a beta channel and a [59:29] deaf channel. And I kind of already set [59:31] things up for that. But we're not quite [59:32] there yet. So So right now it's it's [59:35] it's daily releases. Um, [59:39] and it's very helpful if people say, [59:40] "Oh, this broke right now." It's even [59:42] it's even better if you if you say, [59:44] "This broke right now. Here's a PR how [59:47] we could fix it." [59:48] >> Beautiful. [59:49] >> Um, [59:52] yeah. If you want to help, get on [59:54] Discord. That's the That's the place to [59:57] be. [59:58] >> I'm going to post the uh the link to the [60:00] the Discord invite. It's super simple. [60:02] It's literally discord.com/inviteclaude. [60:06] You can also find it on the website, but [60:07] I'll I'll go ahead and post it on all [60:08] the channels. So, um I've been in it for [60:11] a total of like four hours and people [60:13] are really nice. So, come join come join [60:16] the and funny. They have a really [60:18] fantastic sense of humor which I very [60:20] keen and appreciate. Amazing. Amazing. [60:24] What's uh So, now we know how to help. [60:26] What's like your core priority for the [60:27] project? Like say because this space [60:29] moves in hours, not days or weeks. Let's [60:34] call it end of February maybe. What's [60:39] something that you would like to have [60:40] accomplished by then the project or a [60:42] core priority that you really want to [60:44] hone in? Working on the sandbox and [60:46] security obviously top of mind. Anything [60:48] else? [60:51] >> H you know like it is a it is a one line [60:54] on the website that makes it work. I [60:56] want to make sure it makes it work [60:58] wherever you enter it. Uh which is [61:00] really hard because there's a ton of [61:02] different systems. Um, so like [61:06] it's a little bit of of a never- ending [61:08] game to make sure it it it works, but it [61:11] should be simple to install. I want I [61:13] want not just have this, but I also want [61:15] to have like apps for the iPhone and [61:16] Android and and for the Mac. Um, and [61:19] they're already there. They're just like [61:21] not good yet. So if you want to get [61:23] involved also those are very much blind [61:25] spots right now because I I I started [61:28] building that but now but then the [61:30] project took off and I I had to focus on [61:33] getting core right. So so one thing I [61:37] also want to build is like there's this [61:38] onboarding wizard and it already tells [61:40] you um please read the security document [61:43] because you need to you need to [61:45] understand with a great power comes a [61:47] great responsibility. [61:49] uh for example like not not put in a [61:51] cheap model. [61:53] Um I also want to make it better that [61:57] you can also select okay sandbox and [61:59] like explain people really like all the [62:02] all the stages of how much power they [62:05] want to give their bot. Right now this [62:07] still requires a little bit of [62:08] documentation reading. Um I want to set [62:11] up more people to actually push the [62:15] project forward. I have I just didn't [62:18] find time yet to to think about it more, [62:20] but I want this to not just be me, but [62:24] be more of a community thing. Um, and I [62:29] think they're going to be doing a month. [62:30] >> Amazing. Okay. Well, so I think that's [62:33] something that folks who are watching [62:34] now can help and are keen on helping. [62:37] And for all of you cracked out like um [62:39] iOS and Android developers, like this is [62:42] a fun project. Like go make this app. [62:44] Are you kidding me? That [62:46] >> go make multiple apps. Why? Why [62:48] shouldn't it be a React Native one too, [62:50] you know? [62:50] >> Thank you. Thank you. Yes. I love the [62:52] power of choice that you're giving [62:54] people here. That's like the key. Uh, [62:57] okay. So, those are like right now there [62:59] are some issues that are open. You have [63:00] some PRs that you got to review because [63:02] right now you're the only one doing [63:03] that. But hopefully that will change [63:05] next time we chat. Um, and but I do want [63:08] to ask like I know right now you haven't [63:10] thought about that core philosophy what [63:12] the project's going to be like because [63:13] it's early stages but you've been [63:16] releasing like weekly. Is that a pace [63:18] you're trying to sustain? What? Oh my [63:21] gosh. Okay. [63:22] >> It's it's about two 300 commits every [63:25] day. [63:25] >> Okay. Wow. All right. I guess that [63:27] answers my question. Um, next time we [63:29] talk I hope that you got some kind of [63:31] version in [63:33] and that you're actually sleeping and [63:34] things. Um, [63:35] >> yeah. you know, but it's it's also like [63:39] it it this blew off [63:43] in ways I I couldn't have imagined. And [63:48] I want people to have a good experience. [63:50] Like this is always what's been driving [63:51] me even in my company. It it never was [63:55] about money. It always was about I want [63:58] to make good shape. [63:59] >> Yeah. [64:00] >> And my my my standards for for good are [64:04] quite high. [64:05] Uh, and I know I'm not quite there yet. [64:08] Uh, so so yeah, I'm I'm I'm working on [64:11] getting it to a point where I can be [64:13] really proud of it. I'm already a little [64:14] bit, but there's definitely blind spots. [64:17] Um, a lot of people understand. There's [64:19] some people who like come in there and [64:20] expect, oh, this does not work and this [64:22] does not work and just like start [64:23] bitching. Yeah. But this is not how open [64:25] source works. And also, um, I'm not a [64:29] VCbacked 200 billion dollar company that [64:33] has like an army. Um it's it's it's a [64:36] dude working from home and a lot of lot [64:38] of people that help. [64:39] >> Yep. [64:41] >> Um [64:43] I'm gonna quote I'm gonna quote that. Uh [64:45] that is 100%. We got to remember that. [64:48] So before you go on Discord, you go on [64:50] Twitter and and cloudbot Peter. Remember [64:54] it's literally Peter doing this. It's a [64:56] labor of love. But get involved with the [64:58] community though. Uh would you wrote [65:01] this in Rust or somebody's Yeah. No, [65:03] wait, wait a minute. What is an ad [65:05] saying? Let me see. Seeing the icon, I [65:07] think it's fully written in Rust. I [65:09] think the the the crab has [65:12] >> no more aesthetic [65:14] >> value than than actually [65:16] >> No, this is all this is all typescript. [65:18] >> Type [65:20] I I [65:22] ever since AI, I don't really care so [65:24] much about languages anymore. [65:26] >> It doesn't matter. [65:27] >> It's the language is not what matters so [65:31] much. It's more like the ecosystem that [65:33] matters and and the for this this [65:37] project specifically I wanted to be like [65:38] very approachable and hackable and the [65:42] world's top language for that is [65:44] JavaScript and TypeScript. So plus [65:48] Typescript is really good at web stuff. [65:51] >> Yeah. [65:52] >> And there's this is a lot of web stuff. [65:54] This is a lot of JSON coming in moving [65:56] JSON around pushing JSON out somewhere [65:58] else and then and then go going back. [66:01] So, so that was the obvious choice. I I [66:04] I like building stuff in Rust. I like [66:06] building stuff in Go. A lot of my my CLI [66:08] are are done in Go. Um I play sometimes [66:11] with Sik. I [66:14] obviously like TypeScript for like web [66:17] things. I the the native apps are done [66:20] in Swift because [66:22] the Mac one, the Android is in Cotlin. [66:25] It's it's it's much more about the [66:27] ecosystem these days. And [66:30] yeah, I think it was the right choice [66:31] because it it enabled a lot of people to [66:34] do contributions. It is obviously like [66:37] JavaScript. [66:39] >> Yeah, [66:40] >> it has it has its own baggage. Um, but [66:43] nothing's perfect in life. It's always [66:46] about it's about trade-offs. So, yeah. [66:48] No, [66:50] rewriting that in Rust is probably not [66:52] an option right now. But [66:55] >> it's early date still though. But I I [66:57] appre I appreciate that though that [66:59] cander and actually we all know the real [67:01] language this was written in. Blood, [67:03] sweat and tokens. Many tokens. [67:06] >> Yeah. Yeah. And and many many sleepless [67:09] nights. [67:10] >> And many sleepless nights. [67:12] >> Yeah. Cuz because you know even though [67:14] even though I I mean this project is [67:16] kind of insane because it is 100% [67:18] written with AI. There's like there's [67:21] not a single line in there that I typed [67:24] by hand. But you looked at it, you [67:27] review it. [67:29] >> Let's say let's say most of it. Um [67:32] there's there's there's some code that [67:35] pushes code from one shape into another [67:37] that is just not interesting. Or like I [67:40] don't know the it also has a web server. [67:42] Do I care which class which tailwind [67:45] class is used to align the button? No. [67:48] As long as it looks good. [67:49] >> Yeah. Yeah. [67:50] >> Um but do I care? Do I care about about [67:55] how's the pairing code working for for [67:57] Telegram so other people cannot [67:59] authenticate as me? Yeah. So I look at [68:01] that, you know, it's always like [68:04] you need to have a system understanding. [68:06] You can get you can get away with not [68:09] looking at some stuff. Um but [68:12] you have to look at other stuff. But [68:14] yeah, even though it is all AI, it's [68:15] still a lot of work because because we [68:17] because what those agents still miss is [68:21] vision and taste and love. So, so I, you [68:28] know, there's also this meme often going [68:29] on where you people write a big list of [68:32] things and then they just wal it and the [68:34] agent does everything. [68:35] >> Mhm. [68:36] >> That's not that's not how I think good [68:38] software is being made. like I I I need [68:40] to build something and then I need to [68:42] play with it and see how does how does [68:44] this feel? How does this look like [68:46] actually? And then based on that I learn [68:49] and I adjust my my vision for what it [68:52] is. My vision for what it is is is very [68:54] different than when I started. [68:57] >> Yeah. [68:57] >> Probably probably again will be very [68:59] different in in a months from now as I [69:01] see how how people use it. Right. my my [69:04] my focus on u making it much easier for [69:07] people to also like sandbox and so is [69:09] partly because I see a lot of a lot of [69:12] absolutely non tech folks adopting it um [69:16] which made it to me priority like also [69:18] like made sure they can make great [69:21] default choices but in the beginning I [69:24] put it for me um and clearly I don't [69:28] need any of that but it's it's it's a [69:30] really fun challenge [69:31] >> I love that but you mentioned something [69:33] That's so important and that's I think [69:35] this is how we all know we're going to [69:36] keep our jobs because there is no taste [69:38] yet. I mean eventually maybe one day [69:40] there will be a model that's like wow [69:43] that is really but and especially [69:45] because we are constantly evolving like [69:47] look at you just say how I didn't start [69:49] thinking about sandboxes because that [69:50] wasn't my use case and now you're like [69:52] and now this is what I'm working on [69:54] because I want everybody that might not [69:56] have the technical skill to do it to [69:57] have a good experience. Um, that's part [70:00] of that taste and sensibility and love [70:02] that you can't make that [ __ ] up. Like [70:04] that has to come from a person. So, in [70:07] that note, we're we're our jobs are [70:09] safe. Our jobs are safe. Uh, let me see [70:11] if there are any questions and thank you [70:13] so much. You're being so generous with [70:14] your time. I just looked up and we're [70:16] over an hour now. So, my bad, Peter. I [70:18] know it's late for you. [70:19] >> Thank you. Thank you for for for for [70:22] having me and having me twice basically [70:27] and and for and for for allowing me to [70:30] share this project with the world. Like [70:31] I yeah, I'm blown away by by all the the [70:36] feedback. I I'll keep I'll keep working [70:39] on that. Hopefully building up a a [70:42] community where more people will help um [70:46] and keep having fun. [70:47] >> I love that. Where can people find you? [70:49] And also, is this true that Clawot loves [70:52] to buy things with your credit card? [70:55] >> Honestly, honestly, I I didn't tried it [70:58] yet, but people on Twitter gave it gave [71:01] it access to one password with your [71:03] credit card and bought things for them [71:05] and it works. [71:06] >> Yeah. [71:07] >> I think that the the scariest thing I [71:09] did in in in Marrakesh when it was very [71:11] early, you know, the my very shaky V [71:14] relay, [71:16] I told it to like, "Oh, I need to fly [71:18] home. can you check in my flight? [71:21] >> And it was like sure and was spinning up [71:22] the browser [71:24] and [71:25] >> it could have easily cancelled it. It [71:26] could have easily rebooked you to go to [71:29] Columbia. Who knows? [71:31] >> Yeah, I'm I'm joking. But you know, like [71:33] we had the touring test where people can [71:35] fake as being human. I propo I proposed [71:39] the British Airways login test. [71:41] >> Okay. [71:42] for for for check checking you in on on [71:44] British Airways because it's like a [71:47] 20page form. [71:48] >> They are so worse [71:49] >> on on a on a on a terrible website. [71:53] Uh and one of the challenges was that it [71:55] actually had to enter my passport [71:57] number. So it looked around on my [71:59] computer and it found my passport. PDF [72:02] and I opened it and got the number out. [72:05] And eventually after 20 minutes of me [72:08] sweating [72:10] >> while watching it's like will I will I [72:12] ever come back from Marrakesh we're [72:13] going to be stuck here because my agent [72:15] will like blow up. Uh it actually [72:18] managed to check me in. [72:19] >> That's pretty impressive. [72:21] >> And that's very impressive. [72:23] >> Now we worked a lot on browser [72:24] automation. It works much better. And [72:27] the best part was it it took around in [72:29] the very early version 20 minutes and [72:31] then it bitched about the shadowdom and [72:33] how bad the website is. [72:35] >> I freaking love that. So it's not only [72:37] going to do the job but it's going to be [72:38] very opinionated about how much you suck [72:40] to do. [72:41] >> Well and Logan. Yeah, this is a great [72:43] Rubic. Yeah, [72:44] >> I love that. I love that. That's epic. [72:47] Peter, man, it's been a pleasure to chat [72:50] with you today. I'm gonna I'm gonna spin [72:52] this thing up. I'm like looking around [72:53] like what what what is it going to be? I [72:55] got a old Windows machine I'm looking [72:57] at, but I want the I want the real I [72:59] want the good experience. [73:00] >> Check out Check out we we have semi good [73:03] documentation. Again, I could need help [73:05] to making them better. Um there's guides [73:08] for for Hetera and for some other [73:10] services where you [73:12] >> I don't know it's a few bucks a month [73:13] and you can like have your own little [73:15] little cloud. Um [73:18] or just install it on your laptop if you [73:21] if you want to go all yolo mode. [73:23] >> Yeah, I like this. I like this honestly. [73:26] Um, folks, if you've been building with [73:28] Clawbot, if you've been like it's like [73:30] building your life. There are so many [73:32] things like I can think of a million [73:34] applications where I would use this. I [73:35] love the what you said about it being [73:37] you can have your own independent your [73:39] own agent for your family. Like I sort [73:41] of got to spend like half my life [73:43] reminding people of the things that like [73:46] where we got to go, what we got to do. I [73:49] have a house with two kids in it. You [73:50] know, you know what the future will be? [73:54] Well, like today instead of you pinging [73:56] me, your agent, [74:00] >> no, no, your agent can talk to my agent [74:02] and then my agent can like turn up the [74:04] volume so I I wake up on time. [74:08] >> I love that. I would I would have told [74:09] my agent, what am I going to call him? I [74:12] would have told them I would have sick [74:13] him. Go find Peter now. Go find Peter [74:16] agent and then get agent to to wake I I [74:19] I think somebody posted this tweet this [74:21] meme yesterday. It's like I'm too busy [74:22] right now. Talk to my agent. [74:24] >> I love that. [74:25] >> This [74:25] >> that is the future though. That is the [74:27] future. [74:28] >> This is going to be the thing. Yeah, [74:29] >> I love that. But honestly, to free up [74:31] those cycles like there are so many [74:33] things that I feel like it's been an [74:36] hour and I literally spend an hour [74:38] calling to book a freaking dental [74:39] appointment or to make sure that my kid [74:41] knew that they got to be somewhere. Like [74:43] all those things that cognitive love, I [74:45] want to apply that on things that are [74:46] interesting and exciting. So [74:48] >> even more so you know like after we [74:49] built this one person in the chat wrote [74:52] this is literally changing my life [74:54] because I have so much anxiety in in in [74:57] calling services and now my agent can do [74:59] that for me and I I never I never [75:02] thought of it that way right but like to [75:04] me that that was like a very humbling [75:06] moment. It's like oh wow I or we because [75:09] the original contribution came from me [75:11] we actually changed something that [75:13] changed someone's life for the better. [75:15] This is like that felt good. [75:17] >> Open source at its finest. I love that. [75:20] I love that so much. Thank you, Peter. [75:23] Keep building. Get some sleep. Folks, we [75:25] got to get in this project. I'm gonna [75:27] Listen, I'm going to spin it up. I'm [75:28] going to go through the docks. Uh, make [75:31] a commitment. People in the chat. Tell [75:33] me how you're going to help the project. [75:34] Keep it growing. Peter, you're epic. [75:37] Thank you so much for coming back from [75:38] retirement to make our lives better. [75:41] >> And it was a pleasure talking to you. [75:42] >> All right. Please come back. Okay. Um, [75:45] I'll have my agent ready to go talk to [75:47] to your agent by then. [75:50] All right. Take care. Oh, Peter, where [75:52] can people follow you? Where are you [75:53] most active? Is it said on Twitter? I'll [75:55] share your Twitter here now. I think [75:57] >> Yeah, it's it's it's st on Twitter. And [76:00] I'm also a lot on Discord. [76:03] >> Okay, perfect. And if you want to talk [76:05] to uh Clubbot on Discord, the minute you [76:08] join, it actually quickly starts [76:11] chatting with you, which I I was [76:12] impressed. That was fantastic. Thank you [76:14] so much, Peter. I'll see you hopefully [76:17] in a couple of months. Yeah, you have to [76:18] come back and give us an update. [76:20] >> Do you see this dialogue in the back? [76:22] That's why I couldn't That's why it was [76:23] blocked on the lights. I forgot to press [76:25] allow. [76:26] >> Ah, okay. [76:28] >> Next time, next time the demo will be [76:29] perfect. [76:30] >> All right, perfect. Next time, next time [76:33] we'll we'll have the agents run the show [76:36] here. We'll just stand here and look [76:38] pretty. Thank you, Peter. Appreciate [76:40] you. [76:43] Thank you so much to Peter for your time [76:44] and to the Cloudbot community. Uh legit [76:47] I've been in the Discord server for a [76:49] few hours and it's been delightful. So I [76:51] encourage all of you to go take it for a [76:52] spin. Set up your instances. Check out [76:55] the documentation. If you find those [76:56] holes as you're installing, as you're [76:58] going through your own process, then [76:59] help help with a contribution. So ah you [77:03] guys are amazing. Thanks for the Hey, [77:05] this is the co-pilot hat. I appreciate [77:08] that. Thank you everybody. Have a great [77:10] day. Thanks for coming to open source [77:12] Friday. And speaking of bringing [77:13] intelligence and agentic intelligence to [77:15] your lives, uh GitHub copilot, now the [77:18] CLI has an SDK. It was just released [77:20] yesterday. It is free and open source. [77:23] Uh free, not like free beer, but free [77:24] and like you can uh support and [77:28] contribute to it. So I'm going to leave [77:29] you with a video of that. And if you do [77:33] end up using Clubbot or you do use the [77:35] GitHub SDK, Z compilot CLI SDK to do [77:39] something cool, please tag me. I'm a [77:41] Colombia dev in all places. I'll catch [77:43] you next time. [77:57] Heat. Heat. [78:07] Heat. [78:18] Heat.