Draw in your terminal! tmux + Terminal Maxxing w/ Ben Vinegar

Welcome to Syntax. Today we have the former general manager of Syntax and the co-founder of modem, which is an AI project manager and a nonoing terminal guy. Too often times we have terminal folks who are just why don't you do it this way? Uh Ben has a really awesome setup and we're going to be diving into all of the ways that he is using his terminal for running agents, all of the cool things that he is doing. And man, I'm so excited to catch up with Ben, but I'm also excited to dive into some of his projects because just watching his most recent talk, man, it solves a lot of the problems I'm personally having. So, we're going to dive all into all of that and more. My name is Scott Tolinsky and with me as always is Wes. But most importantly, welcome Ben. Hey guys. It's good to be back. This is your second time on Syntax or the third time. This is the third time. Yeah. Going back a while. But the last one was an emergency fill in. So it's nice to be, you know, properly invited. Sometimes we need emergency fillins and and you're you're just the guy for that. So yeah, you got lots to say. And Ben had a talk at the AI engineer conference a couple weeks ago or and I just watched the video and I thought it was super interesting his like rig of like how he's doing agent coding, how it's done remotely, sshing into boxes and doing it on a box versus locally. T-Max I've been like a I would say it's fair that I'm a T-Max hater. um uh and may maybe changed, you know, never never really understood all of the hype around it, but Ben's usage of it pretty good. So, I think like let's let's dive into that, man. Like not understander. Yes. So, I'm I'm curious. Okay. So, I don't know the background is I'm a Mac enjoyer. Okay. I like ideides. I have used cursor primarily but even before that VS code like probably a lot of people um look even gave the jet brain stack like a pretty good go there for a while. So like that's like I I like debuggers. I like all those tools and everything. And look when we're building modem I think it's worth mentioning that modem is like a uh we started like a year ago and it's AI product management product. It could be project too. That's okay. Oh, product. Did I say project? Yeah, but like you know, the lines blur between them, so it's okay. Really quickly, product management, you know, helping you do the non-coding tasks. I could talk about this later, but um you know, synthesizing feedback from customers, um making sure your ticket backlog makes sense, following up with users when you land some stuff. I'll cut it off there. We started about a year ago and we actually decided that we would codegen the product, like the project from the very beginning. It was kind of like look the whole thesis of modem is AI coding is 100% going to be totally real and if you believe that then it's going to go down like the time to deliver code is just going to keep dropping and then like the other parts of your job are going to be boring and slow. So can we work on that? And that's kind of where modem came from. And if we were going to build it, if we believed in that future, we had to like build it that way from the beginning. So I'd say it's like 99% codegen, but like tasteful curated codegen. Like everything is code reviewed. We go back and forth with the agents a lot. We do edit the code. Man, I was having this problem which was I feel like it's a problem that you're seeing a lot of lately, which is I I don't have a setup where I I have like a there's no Jurro backlog or or linear backlog where agents are just spinning up all day. Have you guys tried that? No. No. I don't I I think that's the craziest thing that people have like agents just constantly picking things up and churning. Like if I if I were to let a rip on my like GitHub issues for all of my projects, I' i'd be done in a couple days and it would not be very good. Yeah, it would not be very good, right? Yeah. I like I've experimented with it. Um even modem because it had like it ingests data and uh like conversation data and says like hey these are like product features that people are interested in. I've experimented with just having like modem spin up agents to build stuff. I don't know. I don't think we're there yet guys. So yeah, I I was just thinking, oh, like how can I get more done? And I started thinking about all the ways that I was being kind of like I don't know, not getting as much productivity as I wanted. And one of the most basic ones, frankly, is that the machine has to be on all the time, you know? And I I feel like you're seeing this online with people joking about cracked laptops or I feel like I even just saw a video of someone with like a harness walking down the street with their laptop like I don't know like this. Oh, I was just telling Wes moments ago I I installed a it's an app called Oh man, what is this thing called? It's called Keep Keeping You Awake, which is like a free version of Caffeinate, which Yeah, it it means you can shut your your lid of your laptop and have your agents still keep running. Even if I had them still running, I don't know. I don't have like a permanent internet. Like maybe you can solve this entirely. I do have a friend who's like walking around with a massive battery in his backpack, for example, and like a permanent internet connection. So, people are doing it this way. Oh my god. Uh yeah. You know, like I'd start a job in the middle of the night like I would I don't know. You ever done this? Like I' I'd cue, you know, like I' I'd write like a plan and then like maybe like midnight I'd be like go, you know? Oh, yes. Yes. And then I come back in the morning and find out well one maybe it got blocked on a tool approval call because I wasn't running dangerous mode. And even even though I thought I'd approved every single tool, like nope, it tried to run one thing that I didn't approve and so it stopped 20 minutes in or it turned off or whatever. So look, I just started running running my agents, you know, by shelling into machines least or whatever, like, you know, get a get a droplet or get a VPS from Herzner Digital Ocean. I've done a lot of that. And I also just have like a pretty beefy machine that I keep uh connected in my basement with uh now three gig fiber, which is pretty sweet. And look, the connectivity one is another one. Like if you're on the plane, if you're on the train, you're trying to work with agents and you barely got an internet connection. As long as you can shell into a machine that has great internet, like you can you can go all day. So anyways, that's the that's the preamble. I'll I'll pause there because I've been talking here, you know. Yeah, I I find that really interesting because I've been finding like my I got the MacBook, M1 just M1 Max, whatever like the the the best one you could buy when when I got it and I've been noticing recently it's like it's starting to slow down and a lot of that is is because I'm running several agents at once. there's background processes, you know, you're running a lot more software, things are starting to get laggy. And I often think like like maybe I should just be running all of this stuff not on my machine, right? You can run it when you're not there. You can run it when like the the whole IBM thin client like are we are we back to that? We we could be. I I'll also add I was hitting the the machine like limits too. Like look, you get two agents that are trying to hit like your test suite at the same time depending on depending on the test suite. Like a lot of our test suite hits the database. So they're pretty heavy. I've hit problems with like max, you know, I have to optimize the number of connections you're opening for example on the test suite, but I get up to I I get up to max CPU pretty fast. Yeah, even on an M5. I don't know if you guys are there on that. Not yet. Not yet. I've been waiting for the next ones to come out. Waiting. Yeah, waiting for the next one. But also it's just like there's so many process that get started and not stopped and just kind of sit around and hang out and take up space and all this stuff. Yeah. So I don't know. Look, I just committed I I may have had a weird setup anyways. Like I really enjoy open code and pi and even just clawed code. I would do this thing with cursor where I'd actually just open up like a window like a dedicated terminal for like cloud code or open code but I would just use cursor for like tab complete and like navigating diffs and editing code. So I felt like you're halfway there, like you're using, you know, once you realize that you're just using the terminal agent quite a lot and like, well, how much do I need this IDE? And so that turned into like, you know, as a norm as a norm core like Mac user, what if I just shell in to my machine and I just do all the work there? You lose a bunch of things. I don't want to say that it like is awesome, but I've learned a bunch of things along the way. So like like tell us about that because like I've I've sshed into boxes before and like what I find is that a it's a little bit limited um because like you don't have all the stuff that you have installed on your machine and certainly you can get all of that set up. So like is that what you're doing? You're just getting like a a box with everything installed that you needed or you just kind of let the the agent install everything. So that's kind of like my first question. And the second one is like doesn't the terminal suck as a UI? I like I use it every single day but like limited isn't it? Well, let me start. I'll I'll give the setup. The setup that I use is tail scale, you know, SSH into the machine. I think you guys have talked about tail scale here. Maybe even a dedicated video if I remember correctly. Yeah. Oh, yeah. I use I love tail scale. I I use it uh for every system. I have even my my wife set up on tail scale to connect to our services. Yeah. Still watching the syntax content. Still following along. Right on. Was there a tail scale episode or did I hallucinate this whole thing? Maybe. No, it was a video. It's a dedicated video. Yeah. Okay. All right. We've we talked about it quite a bit. I've been using Cloudflare tunnels. Scott's been using Tail Scale. So, we've we've mentioned it on several podcast. Yeah, I made I made a video all about my setup, which is similar to yours, but missing several key things that I'm excited to dive into. And then after that, it's T-Max. And T-Max is a window multiplexer. There's other ones of these that are more modern like Zelage. Zelage is a popular one. Played around with that. Um, but I'll I'll explain why T-Mox maybe a little bit more later because there's some real advantages to it. T-Max is a window manager for the terminal. So, that gives you a couple things like one, you can actually create windows um like PES. You can split the terminal just as you could split Ghosty or it iTerm or even Chrome. You can split like a inside a chrome tab you can split that now if you played around with that. So the same thing you can get on T-Max and then you can have like well I can have my coding agent on the left. So I can have open code or pi or claude there on the left and then on the right I'll usually just have maybe just like a raw terminal and I'll go into an editor. I'll look at a diff tool. Um and that's usually most of how I work. If I need to look at a server, I can also get to that via tail scale or often I'll just like build something and push it to a branch and let um I use Verscell a lot like using Versel preview branches do the work for me of like making you know if you have reasonable confident in the agent today to build something somewhat correctly you know then you then I can go and take a look at the preview branch and poke around with it. If we're talking like a web project for example I'm building a lot of terminal stuff so I just use that on the terminal. H how is T-Mux different than splitting your ghosty term? Like because I I use like split pains in my term all the time, but like what what makes T-Max pains different than that? Two major reasons. One is that you can detach and reattach to your team session from anywhere, whether that's Ghosty or iTerm or another computer, you know, in a different room. And I think that's really great because you end up you kind of build a little workspace, you know. So I'll usually have five windows with different projects that I can kind of tab between and there's commands that you learn for doing this and each of those windows has like split panes and sometimes I split them a little bit differently depending on what I'm trying to achieve and you can dis you know so if you disconnect you can reconnect and you get that all back. Um whereas I think if you split all the let's say you do a whole bunch of windows in Ghosty and you split them or whatever if you disconnect you've actually got to reconnect every single one of those if that makes sense. Like they're all dedicated. Yeah. Yeah. Like you're that's just multiple SSH connections, right? Yeah. That's that's no way to live. That's how I live currently. That's that's why I'm I'm so interested in this. Yes. Look, I think if you just want like the UX, I think Zelage Z E L L I J is almost like an easier one to use. I've played around with it. I um some people in this office here that use it. But the second reason why T-Max is this and and not Zelage. T-Max has a CLI interface that's pretty cool. It allows you to inspect active sessions and to send keys and to manipulate them. So from one session you can be like hey go read the logs in t-mucks pane one. Okay. And now your coding agent just has to run a bunch of CLI commands which look something like t-m space capture pane one. T-m space you know send keys one. It can inspect the sessions. It can figure out which pane you're even talking about pretty effectively. So I'm talking about using any coding agent, you know, open code plus like a like, you know, probably like a sonnet or a, you know, even like a Kimmy, like you actually don't need a super smart model to do this. I find like even Kimmy K25 will do this totally effectively. So that's valuable for a bunch of things. One, I don't know T-Max, man. Like I I didn't learn these keys. So, like the first thing I did, and this was a total experiment, was give me a tour of T-Mux to a coding agent. You're in a T-m session. Give me a tour. And it just started opening up PES and Windows and being like, go check this out. I opened up this, I opened up that. And I'm like, wow. Okay. So, you don't even need to learn the commands if you just want to get started with this. You just got to start T-Mux and you tell your agent. They'll probably even figure it out that it's in one and you can ask it what you want. like I want this in this pain on the right. I want this. I want you to split the pain this way. You can just use natural language and you can get the environment that you want without being like an expert. Um does that make sense? Yeah. Yeah. So it's AI you can talk you can control the term directly from your your local agents. Yeah. That that seems very useful to me. Okay. I couldn't get it to work with Zelage. This is one of the things where TeamX is like I don't know it's old, right? It's kind of old. But all that stuff gets baked into the models if it's old enough and popular enough. So they're really good at navigating it. Um, so that's just for like getting you started, right? Or like I still use it cuz like I don't know all the commands. I've gotten better at it. Like eventually you decide to learn the hot keys. Um, but I still reach for the agent to help me out. But I think more meaningful is thinking of T-Max as a superpower where your agent can just go and read these other PES. So I'm running a server in pain two. Okay, I'll go debug that by just reading that pane. And it can even scroll up and down like it has like the buffer and everything. So it can scroll up and down. It can read the log output and go, "Oh, here you go. I've got the answer." Does that make sense? Yeah. Yeah. That Yeah. Yeah. So, I've gotten to the point where I even run T-Mux on my local Mac for this reason. Uh because you can't do it with Ghosty and you can't do it with Iterm. Yeah. Cuz like the the options you have here are like you you need multiple terminal windows, right? So, you're either having the agent spin up child processes which you cannot see or you're opening multiple tabs and you're kind of copy pasting from from one to another. And and both of those experiences kind of suck, right? So it it's nice to if you have to run multiple terminals to use something like T-Max where you can use it like that's in my opinion that's the best AI experience is where the AI can use it and you can use it in in the same um the same way. Like that's why I think the web MCP stuff is going to be so good is that like the way that you use it and the way that the agent uses is the same thing. and and you don't have to have these separate processes which is like oh now I'm doing agent stuff I'm going to do it this different way than how I would normally do it. Yeah and it's open source and it works everywhere. So that's you know you take those properties you hit on something there which maybe this matches for you but like if you use claude you know and it spins up sub agents you can't you can see that the sub agent is like it's really reduced to just like a little like sub agent is working right or you can go and maybe go manually go and inspect it. I think they've got like commands, but you could also just run sub agents or other agents or whatever and just be like use teammucks. And you're right, they they'll just spin up windows or they can spin up windows that you don't see, but they're there in the background that they can then like you could go see them if you want. And I find that pretty useful. I know that there are some extensions for PI, the coding agent, which use sub agents like explicitly through T-Max so that you can see very clearly what they're all doing. So my my question here and maybe we'll get into this, but like the solution to this stuff is just running this terminal thing like like you know what also is is really good at this is a gooey um and like links and buttons and things you can click and and and like URLs that you can you can surf to to see different agents. Like why try to like recreate this whole environment in the terminal? Okay. So, yes, I I am trying to recreate the environment in the terminal and that that was a once you start working this way and you're like, man, T-Mux is good. Agents are good this way. I'm digging it, but now I still have like, well, where's my editor going to be and how do I run the servers and all this? Like, there's still some missing stuff, right? And look, I'm not like a hardcore Linux user. That's the whole point. So, if someone is like going to get really mad at me in the comments, like I'm sorry. I'm sorry if I've gotten this wrong. Um, but like I didn't even know until the last year that you could use your mouse over the terminal. I really did not. Maybe just cuz I had no need. It never crossed my mind that I would ever have to do that. But it does. It works pretty good. It works over SSH. So, not just I think people have seen a lot more Tuy apps, right? Like Cloud Code is the ultimate version of that. it's clear that it uses um the mouse, but it also works over SSH too uh because it's just like um SSH will like pass key codes back and forth that sort of represent where the mouse position is and works pretty well. So you can get um you can get gooies over the terminal. Openy is a library that is what under the it's what's under the hood of open code. Yeah. So if you've used open code and you know that it's got these kind of like user interface mechanisms that you like you can have like sticky components right you can have sidebars you can have uh sticky footers popups toast modals like all of that stuff is actually very achievable and open you can write you can write open apps in a variety of ways I think react and solid so for me as a web developer I was like oh I can just write the things that I understand and I get these like I get these um UI components on the terminal but it's all the stuff that I remember about like you know onclick handlers and and stuff like that does that make sense like render render so one I think you're seeing a wave of like new apps that use the terminal because this technology exists that's cool so a lot of things that I didn't think were possible a year year or two years ago and maybe they always were right it just took me uh took me to look you can you can scroll Cool. You can use your mouse. You can do so much of this stuff. And I use some tools. I could talk about them like um there's a text editor I use called Fresh. Could be like fresh.sh. Maybe you guys can find it. Get fresh.dev. Fresh is somebody trying to build something closer to VS Code in the terminal. So, you know, like look, I use Vim a little bit just because like I'm I'm sure similar to you all like eventually you got to connect to servers, you have to use it. It works everywhere. Yeah. But I'm not, you know, do I really want to edit all my code that way? Not really. So Fresh has um, you know, mouse support and it has um, key bindings that feel more at home. It's also got like what do you call it like a a command pallet that opens up very like VS Code like you can open up I think you can open up a terminal in there like a subterminal um, and you can select with your mouse and delete text. You can go and navigate the menu. Yeah. But you could do all that with a web uh a web UI of that's what I'm trying to understand here is that open code web has like a remote web UI that you got all that and I'm trying to understand it for myself as well is that like I find myself gravitating towards just using the terminal for cloud code even though like the open code desktop app or the cloud desktop app these things are significantly better interfaces. Why do I keep going back to the terminal? Like why why are we trying to recreate uh gooies in the terminal which is like obviously a way worse UI but we like to enjoy it more. What's wrong with us? All right so look this is a very good comment which is many of these tools allow you to just connect to a remote machine and then you can get some of the benefits that way right in terms of being always on. Why don't I do that? That's a good question. one, I've experimented with it and I guess this is probably like they're better now. I want to put a big asterisk. They're better now, but I think when I went back in time when I started down this four or five months ago, I just I don't know. I didn't uh like the experience. The second thing is, man, I use all the coding agents still. I still run claude. I still run codeex and and not being locked down. Like I think that once you if you use open code that way or use cursor that way, you're really committed to using that tool. like you go to the effort of setting it up and you're not really going to let it go. Does that make sense? Yeah. Yeah. No, I I totally feel that especially um like you said, you're changing tools all the time and like Anthropic pulls the plug on using uh your cloud code in in open code and now what do you got to do now? You got to either change the tools that you want to use if you want to continue to use your cloud code max or you're you're going to have to change the subscription, who you're paying, how much you're paying. Yeah, totally. I look I still use actually cursor with the shell extension so that I can navigate the files on the file system. So I still use that occasionally like if I'm really deep into a problem I will actually use some of those capabilities. So I don't want to say that they're not good or useful or whatever but I do like the I guess I like the the portability and the flexibility of just having it this way. A good example is have you have you guys played with the Hermes agent? Bro, currently 15 minutes before starting this call, I I ran the Hermes setup. Like 15 minutes before this, so um I have not, but I am I have like maybe 10 or 12 tabs of stuff open to read after this call. So I'm curious. I'm Hermes Well, look, I I've only spent a couple hours with that, but I would say it's like I didn't have to change anything for my setup, right? I just I shelled into my box. I installed Hermes and I ran it and I didn't have to go and figure out whatever remote magic stuff I got to do. Right. I'm not saying, you know, I By the way, one thing we didn't get into is one of the reasons I make the code work on the box. And um I've also experimented with running VMs on the Mac as well. Could talk about that using a tool called Lima. I want to run everything in dangerous mode. Yeah. Yeah. And so having a box that's just like only loaded with like open source projects and um you know because I can I can you can even segment your boxes to different scoped projects for example you know you can just go ham you don't have to worry about anything. I think that's also quite nice like just the freedom. So tell tell us more about that because I think that's where a lot of people are at right now is where everybody's running claude dangerously skip permissions on the root of their MacBook that has access to like their entire lives which would be like like awful if if there were to be a like a security thing to go down. So like what's your process? Do you have like an image that has a lot of the stuff installed or you simply just have one box and you you SSH in? I think it's worth like look I like experimenting. So I got a bunch of things. I have a framework desktop in my basement on fiber that's running Omari um that I keep up to date and that I pretty much just use for coding and then I think it's got like Steam games on there. it's it's connected to a TV. Like that's it. It doesn't it there's no I'm not signed into anything personal. I guess you could drain my Steam games or something like maybe in the worst case. get my Team Fortress skins or something. Um, but do you not want it to have access to to a lot of your things? Like that's that's why everyone went nuts over open clause and people like I I wanted to have access to my SMS and my my Google calendar. And I guess that's more like personal agent stuff versus like development. Yeah. I Yeah, it is more personal agent stuff. I haven't gone too deep in it. I we've look, we build an agent. Uh I've built a full openclaw style o like open source agent that we run as well on a VPS. That's exposed over Slack and everybody at Modem can use this. So I built some of that as well. I just I accept that people get increasingly comfort with like yolo mode or dangerous mode because 99.99% of the time it's not going to do anything bad, right? And you kind of get you get lulled into this false sense of security. But I've already experienced agents kind of like reason to themselves about why they got to go and do something and seen um seen some bad effects of that. And I just like my position and the position of the company for example is we just everything is gated. We don't just let agents go yolo. A good example is early early in modem's development, we were working with a design partner and what they had asked the agent to do was to create an issue for something that they were looking at. They were using GitHub. They wanted and so modem is a it's a multiplayer agent. It exists on the web in Slack. It's not unlike having like an open claw for your team, but it just does like product and project work. Okay. Uh this is important. It also connects to some of your stuff like GitHub. So we had a customer who our design partner early on who wanted it to open up a ticket for themselves on GitHub. But because of the context, because of what they were talking about, it decided that what the user wanted was to go and open a public issue on an open source repository. So modem just went, you know, as as you've probably seen people talk about, um just went and opened like an open source ticket on behalf of this user, which is not what they wanted. Now, is that going and destroying your computer or like reading your emails or whatever? I think it's just that an agent can it's very possible for an LLM to misconstrue your intent, right? Because if you said to an agent, depending on what agent you use, like, you know, go ruin my day and delete a bunch of stuff, it could do that with an asterisk that Claude will I think has some guards. it'll be like I don't think that you should do that but you know um so if if you can logic if you can like reasonably if the agent can reason itself into a position like it will go and do things that you don't want it to do. I just don't see how that's like unavoidable cuz to me it's almost like it's like talking to a human. It's just pos like what's in your brain and what you want and how you communicate that to the LLM on the other side like something can be lost in translation. It has nothing to do with the effectiveness of the LLM or whatever. Does this make sense? frequently. I would say it almost it often does especially for me because I make assumptions, right? Like and like why would you possibly have thought to do it that way and then you go back and you read your prompt and you're like, "Oh, I you know I there's so much innate things that I think that it it knows when it's really like it's doing what you told it to do and if you didn't give it clearer enough instructions, yeah, very easily get off track." So, I don't know. I I think I'm more sensitive than most and maybe like we're also running a business. We It's like I think that the bar for us has to be higher. I don't want to write one of these tweets about about torching my email. That seems bad. Seems bad for business. Um so to me, I just prefer to put everything in a position where it's like not even possible. Does that make sense? Yeah, totally. I don't know. Cautionary tales, you know? So maybe that's Yeah. How about you guys? do uh yolo mode on the on the Mac. I was going to say, have you considered that having one of those blog posts would actually be, you know, all all news is uh good news. All all publicity is good publicity. You know, you could get your name out there. We [ __ ] up our whole or we deleted our whole database and now uh everybody knows your company name. I think that it depends on what you're doing. Like for us, we ingest like sort of product conversations that could be happening on Discord, public GitHub issues, it could be support tickets. I think that we want to demonstrate a high bar towards privacy and security, right? And I I just don't think that it serves us well to do that. Now, you could still be right, Scott. You could still be totally right. Like, even though I want to conduct myself that way, maybe the right move is to just like drop the database. Do you think we'll get to a spot where the models will be able to figure out what is okay to run and and what is not okay to run? Um because like everyone's everyone's just dangerously yolo mode right now and we hear every now and then we hear these these stories and whatever and like the the only the alternative is like do something like you're doing which is like set up a box but like eventually you're going to give it access to something that it needs and it like you could like at some point you need to give it access to a database and that might be a prod database and and therefore it should be able to to drop the entire thing, Right. So you you sort of just slowly add in more access to this thing which makes it a little bit more unsafe. It doesn't access to your your whole computer, but it does still have to have access to your GitHub and maybe your email and maybe your your actual database. So that that's a problem that could happen there. Um, but the other option is just like you have to sit there and approve approve approve every single like ls-l and there's there's got to be something in between where the agent would be like know like oh yes I can run a listing of files but I shouldn't be able to nuke 18 gigs from from a database. And I'm curious if you think we'll ever get to that spot. Man, that's a good question. I feel like it's it let's say the a let's say the coding harnesses or the LLMs have more of these like protections in place. The thing is how does it know when you want to do something destructive? That could be valid too, right? And a good ex I do want to mention this. I'm I'm using Linux on the desktop and I'm using this stuff. I use the agent to like I could not use Linux without coding agents. I'm sorry. I need a coding agent to be like, "Why is the resolution half? I have no idea." Oh, same. Yeah, same. Okay. Why the heck can't I uh use uh SMB? Okay. You don't have it installed. Okay. Let me Yeah. Um I do, if you heard, I use Steam. A little sidetrack here, but I think this is fun because I use Steam. Uh I use this framework desktop on Linux as like a gaming machine hooked up to a TV. And okay, yeah, it works because of all the work that Steam has done with um this like compatibility layer Proton for like the Steam Deck. Um I had one for six months, didn't use it, but it works on Linux desktop. I had a game that by all measures is very popular and should have run, but I couldn't figure it out at all. So, you know, coding agent, why isn't this game running? and it just started ripping looking at all the crash logs, you know, and after five minutes it's like, "Oh, you got to run with this command and that should pave over this compatibility thing." Man, I googled this for like an hour, you know? I just could not find anybody with the answer. I was shocked. I just want to make a comment that like coding agents make you doing stuff on Linux way more approachable. Totally. But at the same time, it's going and messing in, you know, it's like, "Oh, let me fix this. I'm just going to go rip and like write these system files and stuff. Okay. And I want it to do that because that's part of what makes like that's part of what makes it really powerful. And so I'm I'm connecting this to say sometimes you want it to be powerful. Sometimes you want it to arguably be destructive. And how do you know when you want it to do one and when when the other one is right? I don't know. I don't know. Does that make sense? Yeah. Totally. No. Totally. Um, tell us some more about the tools that you've written here to make it. You've got hunk and term draw which help you take the I guess the asky experience and make it a little bit more gooey. Yeah. So, I mentioned like fresh as like a an editor that I think just makes a little nicer to use on the terminal. I realized maybe like two months ago that I was suffering from AI psychosis. Oh, yeah. And I felt like a lot of people were two months ago because maybe like 46 Opus 46 has come out and we were all having a lot of fun generating lots of stuff and I realized that oh no I have not been paying as much attention to the code as I have been and uh specifically this is not for like modem but I meant for like all my little open source thingies you know I was just kind of like I was vibing more than I'd ever realized part of it was I just didn't really have a nice way to like look at the diffs frankly because of this like way that I had adopted like you know I was using the I was using the stock diff program and then I started using looking at different um diff outputs for git like that that connect to like connect to git like I don't know if you've played with some of these but delta is one of them diffastic is one that's popular it does like a more like syntax aare diff which is kind of nice um these are all like things that you can plug into um git diffs pager you know diff pager mode today and you can do it on Mac you can do it on Linux um and if you do use git on the terminal I use get on the I like for some reason I never let go of get on the terminal so so that's where I started but I was just never satisfied for a variety of reasons one is I just didn't like their syntax highlighting choices I also a big one was they're not responsive you know we're so used to responsive layout s on the web, But terminal programs, not so much. And you know, sometimes depending on how you're working, I've got like a really I've got a split pane. I've got a really tiny surface area. I don't want split diffs. I want them to be stacked, but sometimes I've got like a really wide area and I want them to be split. And I couldn't actually find like a tool that made that really nice. So it just wasn't nice. So I wasn't I wasn't reading the code because it wasn't that nice. And then second of all, I was really interested in some of this new technology that had come out. Open mentioned earlier. Um, and then Pierre diffs diffs.com. We had them on the podcast last week. It hasn't, at the time of recording, we have it hasn't even gone live yet. Was that Jacob? Who you got on there? Uh, other fellow. Yeah, we had Alex and Amadeas on. Okay, great timing. So, like that is a great library and you're seeing it pop up everywhere like anywhere. It's a great diff rendering library. I guess is going to be on the show, but I understand that they were basically building a GitHub clone. I actually use that product and to use modem in the early days. I could talk about that if you're curious. But so I was already very familiar with the whole like their whole approach to diffs and they put it out as a library and it's meant for like it's web tech right primarily. But I was curious if you could plug it into Open Tui, right? OpenUI supports React. Diffs.com has a React output mode. And I was like, man, is it even possible that I could plug these two things together? We were under like, how the hell is he doing this? Yeah. So, that's the And then, would it even make sense? Would it even look good? And the answer was yes. Um, I had to do a lot of optimization. Like, I've used the agents in like, you know, goal mode or auto research mode to like get the frame rate up. And I had people before I released Hong, people were like, it's a terminal app. Why do you care that like about the frame rate? And the answer is yes. does matter. So, so that's where hunk came from. So, you can check it out at um you know, I don't know, show notes, but um github.commodemunk. Oh, you need a cool domain name for this. Take a page from the Pierre guys. You need like We have modem or sorry, we Yeah, we have hunk.dev. I I just got it. I just haven't put anything up there. Um you have got hunk what? Oh. Hunk. That should redirect to my website, honestly. Oh yeah, it's a fun name too. I think that's a fun name. It's a fun name for sure. So anyways, it kind of works. And then after that, like this came out like two months ago. Like I was using it quite a lot. And I think that this is always like a nice thing is like when you're using software and you're using it a lot, right? I know you guys work on stuff to maintain the pod, right? Um you just keep it just gets better. So you know, like the diffs look nice and I put in some themes like it just looks nice, I think, relative to most of the solutions on the terminal. And it has mouse support. It has a sidebar with a file picker and that's gone through like a bunch of iterations. It's got some other interesting features. Probably the most notable one is it has a CLI totally ripped off from T-Max which is that your agent can write and read to it and it can navigate the session for you. So um it takes like rev ranges. So you can say like hunk diff main. Like you can diff main, you can diff, you know, main at five, you can do like you can do a diff on a branch or you can actually just ask the agent to drive it just like you could T-Max. Um and it can leave annotations. So it's a more not only is it I think like a kind of like a really nice diff viewer, it's also got this kind of agent integration via the CLI which I think is kind of neat. So check it out. hunk.dev. We'll have that live by the time this goes out. Um, I can't believe you did hunk.dev when hunk.rodio was available. I Hunk Rodeo. Hunk. Let's get that, too. Need to Yeah, it's under 20 bucks. We'll get it. What does a domain cost? 116 bucks for the first year. Ah, no thanks. No thanks. Yeah, also, you know, questionable content when you Google Google this thing. Look, I don't know. I was making this and I was sharing it. Um, but not like going too hard. I was using it a lot, but it's actually it hunks seem to get a lot of attention lately. And I guess at this point, like I've even forgotten why it's why it's good, but people seem to think that it's good. I don't know. I think part of it is that look, I I really actually liked GitHub pull request reviews. So, it's taking like a lot of that UI if you're familiar with GitHub, you know, code review, you know, you want to leave a comment back to your agent. You want to um navigate between different files. It's got a different it's got a similar kind of like file navigator. So, that that's what's going on there. And for me personally, it has helped me review more of the code that's happening on this remote box over SSH. You got me googling right now. Can I install Omari on my I've got like a a 2009 Mac Pro right now. If it's Intel, you can. It is does not do well with the M processors. There's like a a 4 because I looked into this because I was running it on my old system 76 laptop and I'm like I would like this so much better on a real machine on a real machine and uh I don't have it. I've been running like I have an old MacBook Intel MacBook and I had my OpenClaw running on it and I the other day I haven't haven't messaged my poor little Open Claw in probably 3 months and I just messaged it the other day and it's still running like it's laying on the ground three months later like that's that's amazing. Um but I I do like that workflow. So you got me you got the gears turning here. Anyways, that's just a bit of a side that let's hear about your uh your other package that you've you're working on. So that's one and this is in pursuit of a you know my pure portability ID pseudo ID terminal setup. The other one is term draw. I started experimenting experimenting with writing ASKI to agents quite a bit maybe because I I started as a very young person on bulletin boards before I had the internet like I had a dialup modem you know I'm not that old but like just imagine like a 13-year-old you know connecting to computers that he should not be um in the 90s. uh like the internet existed but you could connect to like a local bulletin board and so to me I I really have this like headsp space of asky and anie uh character drawing so ignoring the terminal like if I was using cursor or conductor or these other tools that I've tried have you ever had this experience with an agent where you're like gh all right you built this web thing but like I look can you take that button you can put it to the left you ever done that and then and then it's like oh you mean over here and it like puts it in a in another location you're not happy with and then you're like no no like can you drop it a little bit and then maybe you'll share a picture like maybe maybe you'll take a screenshot and you'll like draw an arrow about like what you want. Have you done that? I I I thought this was kind of interesting because you said in your talk you said like like why give it an image when it all it does is it converts it to text and when you can you can just give it a thing. And I thought like I actually like I find screenshots work almost better in some cases than like hooking it up to like a Chrome Dev Tools MCP. Oh, they totally which is amazing. But like like the I guess Unreal that visualness is better than the actual raw data of how something works. So is that true for diagrams as well even if they're done in ASKI? My experience has been so I guess I started experimenting with this where I would just draw type it out. I would just like literally type out asky little diagrams and be like look I want the button here. And I found that to be pretty successful and this is going back you know to December. So I was already down this path and then at some point I'm like why am I typing this out? I could just draw it. So term draw is like a a terminal illustrator. Okay you can draw shapes. You can draw boxes. You can put text. Um, you can you can write text. It's got now elbow connectors. I think I just added you can have when you're drawing boxes, you can use different borders. You can put like a dashed border to indicate something's temporary. And I say it's an illustrator because these are actually like objects that exist on the on the pane or like on the drawing surface. So once you draw a box, you can move it around. It's got bakedin grouping. So if you want to if you if you draw like a window and you want to do like hey I want a two column layout you can resize things and they stay in they stay in place. So it has some context about oh I'm a box inside of another box. So if you want to drag things around so it's got some kind of like nice things. You can you can grab the corners and resize stuff. Um and even the text is like movable. So I've done all this basically so I could just like diagram out things that I want um without having to explain it. Like I can open up term draw. I can draw what I want and be like go and do this, you know, in in 20 seconds versus having to type out, you know, well, I want a website with two columns and I want on the left side I want this, right? Instead, you can just really draw it out. And my experience has been that the agent totally reads this quite well. I've done a bunch of experiments. Opus 47 really good at really good at asky diagrams. If you want to play at this, I feel like it really leveled up something else. I do I noticed that too. Yeah, I like it when I'm doing like a grill session and it's giving me asky diagrams. This or that. This or that. Yeah, I Yeah, I I experiment with having um I've been experimenting also with having the agent like draw asy diagrams in hunk annotated comments, which I think is fun. So, you know, leave a comment, but it's like I'm going to write like a sort of layman diagram to help you understand what you're looking at that I wouldn't necessarily want to commit to the code, but I'll put it here as an annotation for you as you're reviewing it. So anyways, that's term draw. It helps if you're I'll say it's also helps if you're if you're doing terminal apps. If you're building terminal apps, it's pretty useful to draw terminal terminal constructs. That's wild. It's just like a whole guey. But and the the format it saves in is like a TD.json file. What is in that file? Is it literally just the the asky as text or is it like more structured? Oh, that's like so normally it just outputs to standard out, but somebody added added that is like, well, I want to save it and come back. Oh, I see. This is another topic, but I'm having a struggle understanding whether like feature requests are real or not. Okay. Which is like so these these projects have gotten like some minor notoriety and now I guess I'm stuck in the world of like people are opening up poll requests and issues and understanding if what people want is real or not is becoming hard. That's just a side topic. And it's is it because like you don't know if these are real people or do you think it's just somebody that says like this should be this is possible because I can type my any thought I have into a box and and code it up. Yeah, man. So many like I somebody on hunk for example, I'm like man, how is this person opening up so many poll requests and why are they so on point? And then I realize that it's like they're generating a PR for every single issue automatically that somebody opens. Ah, and then I'm like, you know, I thought you were just a user who was really into this. And I go and look at their profile and they've got like, you know, 10,000 commits all from the last few months and they've contributed to like 160 different repositories. And I'm like, oh, but like who's who's paying for that? I know. Yeah. I don't know. Is it is it people that have these? Like often I find it's it's people that have like a startup in the space and are trying to prove that their thing is the best and then they're just spamming everyone. Is that what it is? Do you think in this case? Like, you know, I don't want to name names. I think someone could easily find this, but it didn't seem that way. It actually just seemed like they were taking a lot of pleasure in contributing to open source as far as I could tell. Maybe they work for Amazon and they're trying to get their token count up and they need to burn them on something. Or maybe they are building something behind the scenes and maybe they're not ready to reveal it yet, but you know, here's one. What is my merge accept rate, right? You you can actually get some you can get some data on this. Interesting. Crazy. So anyways, yeah, the the format is just if you want to save it and come back. I don't know how many how often people are doing that, but the person who requested it when I didn't merge, they came back and said, "Hey, I really wanted I'm waiting for this." And that's a good indicator for me that maybe they're real. Can I ask what your your flow is for things that must be visual? Like obviously we can make diffs and and diagrams and everything in the terminal, but eventually you need to view a video, view a photo. Um, and I know that there's like some terminals that can display images, but like like what happens if you just want to like look at a PNG image and like what what's your process for that? Well, if you're using Ghosty, which is very popular, it uses Kitty. Kitty. Yeah, thanks. I'm like, what can I say that won't have uh Linux people yell at me? I think this is fun, too, which is like um the protocols have changed. Kitty is uh terminal protocol and I think I think it works over SSH and it works with Ghosty and you can actually just render PGs and graphics. Actually my talk at AI engineer Miami that whole presentation was done on the terminal even though it has full images is done over the terminal over SSH so not just a local terminal I even experimented technically you can you can play video over the terminal however it's super inefficient and I wouldn't recommend it it's really just flashing images quickly so like yeah so that's one so just we're talking just talking images and that's a cool one otherwise I think it's like the standard thing where you expose you use tail scale you expose a port I connect to a server you know you could still do that right gets a little more painful I try not to do that if I can I guess another one is just building you know using build servers and generating previews like using versell there quite a bit but that is like costs money some compute right and it's like it it's like slow as well like I I was using the like cursor cursor cloud agents and they will spin it up and they have like a browser built in to their like cloud agents, which I really like because you can just fire something off. You can go on your phone and then like if it's not possible that way, some people are like, "Yeah, just send in a poll request, wait for it to build." But like that's kind of slow as well. Like I feel like there's there's no like killer solution there just yet. And look, if I was building an Electron app or something, this probably wouldn't be a very good environment for that, Right. So I think it depends on what you're building. Um maybe that that's a a good transition. and we got a couple minutes left here is that you have a podcast um about a dentic coding um and then you've you've developed some skills to edit the thing like what's your process for that and where does that happen? I guess we're plugging or the plug-in can come at the end or it can come now. you plug whenever you want. Yeah, there's no rules. It's called State of Agentic Coding with me and Arma. took us probably three episodes before we committed to that name and it's just once a month and um you know we're just talking I think the episodes are like 90 minutes now look I learned so much working with you all on syntax actually we so loved working with Randy super producer Randy and I hear that he's gone pretty deep writing code and stuff too oh yeah we're uh he's he's he's writing his own replacement because he's he's he's a cracked claw engineer Yeah. Now that Da Vinci has MCP and all all kinds of ograph support. Yeah. Yeah. Da Vinci MCP. He's like he built like a whole like little chat window for it and he like Oh my god. And we were like fighting over like what the best way to do graphics was and like we think like Randy might have cracked the best programmatic graphics approach. I think he could run a course right now and I would attend it. Um. Yeah. So, we actually worked with Look, one of the things I learned is how great it is to work with a with a professional editor. So, we hired um an editor, editor Sam, who um was actually somebody I met. He was a candidate for Ry's job. Okay. And I just kind of stayed in touch and um you know, I was like, "Hey, you want to work on this project?" So, work with Sam. But I think for me, like there's still so much work. You can't just like have an editor, right? It's one thing with Randy knows he's worked with you all for so long. He knows how you all, you know, what what you're going for. Um, in our case, it's a little different. I need to watch the video. I need to come like what do I think is the best, you know, tasteful bits. How do we cut a bunch? Like obviously Sam can cut a bunch, but we're also trying to do like a really fast turnaround time on this because AI moves so fast. So, if we can just accelerate that, like we try to do like four days and then Sam gets me a cut. I got to go and review it again. I'm very nervous about saying something really stupid, guys. Um, I'm really nervous. So, I built a bunch of tools to help me with that. Actually, it's it's open source. I just open sourced it in case people want to play with it. It's called Pod Guy. Can link to it. What's interesting about it is it's um it's a bunch of extensions and skills. Okay, I think people have seen that. One of those, for example, is it'll download like a whisper model and like generate a transcript, but the skill file knows how to do it. Another one is like I've, you know, find the boring parts to cut in this 2-hour episode. We might have two-hour episode that cuts down to 90 minutes, right? Just make some suggestions. Um, one that's actually very helpful is by the end of the edit, Sam has inserted all these interstitials all over the place and uh, I just find it easier to be like, "Okay, pod guy, just find me where all the interstitials are so that I can time code those for the YouTube description or whatever." And then it just goes and like actually generates a bunch of frames and scans through them to figure out where they are and then like generates, you know, goes back to the transcript and generates like all these like descriptions for me. So that's just like it does a bunch more. I think what's a little bit more interesting about it is it's the first time I've used Pi. Pi is loaded in the repository and when you fire up the project, it's actually like a slim down pie where it actually kicks out all of the like it shouldn't it doesn't bring in your skills or your extensions from your mo your local development or whatever. It's it almost only exists in this world to serve you and the podcast editing role. It's like a a dev dependency of the project. Yes. And it's almost like an alternative CLI where instead of you trying to remember like all the commands which I'm never going to remember, it's very much like here's the eight things I can do. What do you want me to do? It's not there to write code. It's just there to actually like almost like it's an agent dedicated to doing this. I've been finding that as well. like I've replaced so much of my like script writing and like these like rigid npm run whatever. I've replaced so much of that with simply just like bunch of skills and an agent and saying like this is kind of what I want and it's just so much more flexible and it's it's it's not like a it's not writing code or anything. It's just like an easier way to have a CLI. Yeah. Uh I think Pi is a really interesting one for this because it's I don't know. Um, have you guys had Mario on here? Two of them. Yeah. Yeah, I did see it. Package deal. Yeah. Because Pi is so stripped down by default, right? It actually works really well for this purpose. If you want to like stand it up as your interface to whatever versus say, you know, Claude, you could do this with Claude, but Claude has a bunch of other things that it wants to do. And it's nice to have something that's just sort of like bespoke tailored, you know, for for this very specific task. And I guess to be ultra clear here, it's not like I run the coding agent. It's like I literally Yeah, I run like npn rundev and that brings up the coding agent. that's great. All right, let's let's talk about modem. Um, so obviously you're a company. Tell us we we touched on what it was, but like tell us what it does and then like like why should we care about something like this versus just having cloud code do it? Okay, so modem is two parts. One is it's like a product intelligence platform where it integrates into all these places where you and your customers could be having conversations about your product like support tickets, public community, Discord, Slack connect channels. Okay. And those messages go into basically like a a pipeline that's not unlike sentries. We're very sentry pilled and um basically just kind of creates like a you know we call them topics but they're kind of similar to sentry issues like instead of you instead of getting a list of like all your errors that are grouped together we give you a list of all your all your kind of product topics grouped together right here's all the places that people are talking about dark mode or you know what you just launched version 15 of your software and starting you know an hour afterwards you had this big furer of people complaining about this aspect of the product like oh you broke mouse support in version 15 right and so all of this is streaming into a platform where it's categorizing it organizing it dduplicating it so that you and your product team can just be like man what do people think about our software right now and you just have everything there and as new you know you launch new things or you have you introduce bugs um if people are talking about those bugs it'll wind up into the platform. Does that kind of make sense? Yeah. Optimum RX, sign up for this. Optum RX, I'm begging you. Uh Wes and I had an epic rant on the Sounds like it's my biggest enemy right now on the web. Yes. Optum RX. Is this a drug plan? Yeah, they it's it's like subscriptionbased. Your meds get sent to your door instead of having to go to a pharmacy. but uh they like rewrote their platform and it's total dog. Uh so often you're visiting a website and you're just infuriated about it or there's a part of a product that you use every single day that you just feel like the people that are working on the product have no idea that you're so frustrated with it. And I often want to be like just I'm going to get a job there, fix this, and then quit. Like I just like I feel like I know better about the problem with this piece of software than anybody at this entire company does. And I wish that I could just get that across somehow. But in in the olden days, you had to get someone to care about about your actual problem. But what if you could just simply take everybody's frustrations and complaints and all of these signals that something is wrong with your product and and like use modem, right? That is exactly it. You sold it way better than me. Thank god. um which is you know like even at century like we had this you obviously we have great observability into the performance or the errors right but what do people think about it I don't you know usually you need the CEO to come you know the CEO basically drives by and sends you a message because he's he's scanning all the channels manually that's like that's one thing that David does I wanted to know I I wanted to know what that is like or sorry if I were at Optum RX I'd want to know that right but often it gets bottlenecked behind and people, tools, services, etc. the engineers, the software developers, the people who probably care. I bet there's people over there who care, but they can't even see it. So, we're trying to like connect that feedback directly to the like software teams. So, that's like part one. The second part is it's connected with an agent that is not unlike you can call it you know claude or you know linear has an agent where it's like really um it's built on that data and it's a multiplayer agent that you can work with over the web or in your slack channel in the future discord sorry in your in your Slack workspace that is really just like someone that you can one you could talk to it to just extract these answers really quickly because it's designed to work with this data really effectively. So, hey, what does everybody think about our new RX platform? Give you that answer really quickly. But it's not designed to just be like a question box thing. It's like it can take action. It can um it has automations. It can go and curate your backlog automatically on your ticket tracker. It could actually like surface some of these problems directly to your team on Slack. Um, so the idea here is it's we're trying to get to like a a vertical product agent, PM, project manager, product manager that's honestly like working with your team and trying to surface this stuff directly to you, you know, cuz like because we don't want another tool that things get lost in. So, heck yeah. Yeah, I love that. Now, you got to scan YouTube and uh podcast for sentiment about products. This is a thing we get to dog food it a lot. like um a lot of people are using Hunk now and opening and like leaving messages on social or opening tickets and I'm like oh man I actually I really need you know uh on top of the fact that we dog food this for modem I also now need to dog food this for our open source projects and like it's it's a lot of work to keep on top of this. So, yeah, man. Do you do you ever see something like this being applied to stuff that isn't software? Like I'm just thinking about like like a physical product or even just like man like Lululemon has been getting like tons of hate lately of the their designs are not good. Um their their fabric is peeling just stuff like that. And I'm curious like does Lululemon have some mechanism to figure out what are people saying about our latest leggings? I think on the conversations I've had is that many companies have something like this like everyone understands that it's a problem they got to solve and they and they have this kind of visibility problem. So I I talk to companies all the time which just like gh we built a version of this there's one person who maintains it and it's not that good. Um so I don't know what the answer is like they might have something like it might matter to them. Um, yeah, but whether it's good or or or you know, not good or doesn't exist, like I think our goal is like can we just make the best version of that and then you know based on working with customers and design partners, can we just make the best you know give everyone like a great version of that so they can uh respond quickly. Sick. Well, I love that and I I can't wait to see this in action more. Um, and everybody who wants to check it out, it's at modem.dev. We'll have the link in the show notes for you to check that out. Uh, if I I will say even if this is not a product that you want to use, go to their website cuz it's a cute looking website. I love your website. I love your colors. I love your animations. I love the whole aesthetic. Uh, it looks great. We didn't vibe generate it. I believe that. I can tell. Yeah. So, that was awesome. Thank you so much, Ben. Uh, we've already done quite a bit of plugging. So, do you have a sick pick? I know you're familiar with sick picks. Uh, did you bring one today for us? There's a lot of chatter about taste and you know, Sel Pie from Cloudflare and Party Kid has talked about like, look, if you want to get taste, just go out, go to museums, go watch movies, just watch a crap ton of movies, listen to music. And I do agree and I believe with that. So, I'm gonna pitch a movie. And that movie is Nirvana, the band, the show, the movie. It's sick. Okay. I guess it's an indie film, but it had like a $2 million budget, so not that indie. Um, it's a Canadian movie. These guys used to have a web series turned into a TV show on Viceland. And it's basically about two guys who have a fake band called Nirvana. And whether it's a real band or not, it doesn't matter. But every show I guess is like they are just trying to get a show. They're just trying to get a show at the Rivlly which is like a a concert venue and it's just like a hairbrain schemed. I have never watched this show before. Okay. I have no affinity. I don't I like I'm not connected to this at all. But the movie is really good. It blends like what's real and what's not. If you like like Nathan Fielder content, you'll probably enjoy this. And it's also a time loop movie. Which it's, you know, I'm really into time loop movies. I won't give too much away. Do you see Palm Springs? That was a good time loop movie. Great movie. Yeah, I think time loop movies are good. I actually I'm now on back on time loop. So, I actually also just bought a copy of this movie here. Two sick picks. Um, called Time Crimes. I got it. I got I saw it on 4K at the uh movie the at the local bespoke movie store. And this is a Spanish Time Loop movie movie from the late 2000s. Also really good. Go check it out. Time loop movies. Well, thanks so much, uh, Ben. I the moment I saw your talk, I was like, I cannot wait to dive in. I already have T-Mux Ripping and I'm going to figure it out. So, I'm gonna ask the agent. Awesome. That's I'm gonna give it another shot. Appreciate your time, Ben. And uh we'll catch you on the next one. Always a pleasure. Peace.