Getting started with Grok Bot

0:00Hey Crockbot, could you go get that information off of Twitter for each of my followers either through the API or by physically going and looking at their profile, create a notion database, pull that information into the database, um, and then hand it over to me. Now, typically this is something we might use AI to write a script for to accomplish that goal, right? This is typically like a personal scripting or software approach. Now, I just had an agent do it and it's kind of better because it's completely disposable. Um, it took me maybe five or 10 minutes and now I have access to this agent forever. I can

0:34tweak it. I can update it and I can approve it. Today, we're going to talk all about Grockbot. Specifically, we'll talk about what Rockbot is and how it works. And then we're going to move into how you can get started and how you can set up the tool to start automating your own workflows. And finally, we're going to dig a bit deeper and I'll show you some of my favorite workflows and uses for this tool because really uh the possibilities are endless, as cliche as that might sound. Um, and we're all still kind of figuring out what this tool is capable of. So, what is Grockbot? It might be easier to show you

1:03rather than explain. I have a bot open in front of me here. Under the hood, Grockbot is just an agent with a computer. So, an agent, as you're familiar with, um, is just a model that has access to some tools and can do stuff for us. And this agent has access to a computer. So if I click this window, we can actually see the computer. More than that, we can

1:23interact with it like a Ramount desktop. So behind the scenes, this is running on a virtual machine in the cloud somewhere. So none of this is happening on my laptop. And I have a terminal in front of me. I could run a command. Um I have a file system and I have uh X here. Now, the cool thing is that Grockbot has access to all of these things the same way we do. So, more than just an agent with a computer, it's an agent with a

1:50computer with really good computer use. And we can actually teach Grockbot how to do certain things on this computer by recording and teaching a task to Grockbot or by prompting and instructing it. And so today I'm going to go over some workflows and some things that I found Grockbot to be really good at. But just to give you a bit of a taste there, that might include things like doing grocery shopping. For example, I asked Grockbot to look at Costco and look at Amazon and add certain things to my cart here and then compare prices, um, compare delivery, uh, charges, compare, um, times, etc. And it's really good at

2:30stuff like that. We can control the computer just like Grockbot can. So when it needs us to step in to complete two-factor authentication or log into a website, it can hand that computer off to us and allow us to complete that action. So recap, what is Grockbot? It is an agent with a computer that runs in the cloud. So nothing's running on your machine and that gives us access via mobile and via desktop that has really

2:51good computer use. And that's really it. That's the foundation. So while the inputs and outputs are chat, we're always typing text. Um those are pretty simple. Under the hood, there's this whole like operating system you can think of like as a server or a really powerful tool that allows us to get stuff done. And that's the foundation for the rest of this video. So, just to drive this home, I can come over to my marketplace bot here and ask it to do a search uh for a variegated monster elbow. Um, and what we'll see in the computer use on the right, uh, is that the bot, because I've already authorized

3:28and logged into Facebook, is going to be able to open up that machine, uh, go to a new tab or maybe use the same one and do a new search. It's going to take all the actions that we would take. So, really, it's more than a personal assistant. It's more than some of the

3:42other AI tools you're familiar with. It's an agent that can do everything that you can do on a computer. So, we can see Grockbot performing searches for a few different plants um and only finding a few of them on Craigslist. So, what can Grockbot do? I want to give you a few use cases before we dig into how to use Grockbot and how to get things set up so you kind of understand why you might adopt a tool like this if you're still curious. I've been using Grockmot for more main four main things which is first writing software. Um second

4:11building personal software as agents. This is a little bit different. I'm actually moving the software into chat. We'll talk about what that means later on in the video um for information and knowledge retrieval. So looking across the cursor and SpaceX, Slacks, notion, GitHub, it's been really helpful. It's actually how I onboarded to cursor. We'll use that as an example. And fourth for automating workflows. So, very similar to what I just showed you with this grocery example, you can think of any workflow, any sort of knowledge work you might have to do, whether that's going on LinkedIn and reaching out to people or finding information if you're

4:45a salesperson, bring that into your CRM, reaching out to these clients, drafting emails, etc. A lot of our go to market team actually does just that with Grockbot. Similarly, for our most recent launch, I actually used Grockbot to coordinate across all the different feedbacks receiving for our launch in Slack in notion and then to sync all the content between notion to PRs to incorporate notion comments, PR comments, Slack feedback, and then to kick off cursor cloud agents to make sure that all the feedback was integrated and it saved me a tremendous amount of time and it was a really great use case. So, we'll dig into those four

5:22things, but first I want to show you how to set things up. So, what are these agents and how do they work? Well, let's take a look at an example. And this is one of my favorite, my tech demos example. Um, and if I open up the settings here inside the bot, we can see there's a name, a title, um, and a description. The name is technically um, optional and doesn't actually influence the behavior of the bot. The title is a description of what the bot does. And then the actual description is a longer form um explanation. So this bot looks at my Twitter bookmarks. I call it like

5:57uh colloquially the Devril bot. And each day um comes up with something that's cool out of those bookmarks. From there, it drafts up a prompt in my style, sends it over to me, and then I can approve or deny that prompt. Behind the scenes, it's going to take that and kick it off to a cursor cloud agent through the cursor integration. And that brings me to another really powerful part of Grockbot which is that it supports all of the integrations MCP servers and plugins that cursor does as well as skills etc. So we can add things like Gmail, Google calendar, Google drive um some of the more complex ones and

6:32internal things that our team are using here as well as other SASes like Amplitude for data analytics, Twitter for accessing uh more structured data, one password etc. Uh and the nice thing is that we can even add multiple accounts. So I have multiple Gmail accounts there. If I go to um actually notion, which I think is the better example. I can have my SpaceX cursor and personal accounts. We can retrieve information across all those sources. So this has cursor as a connection. And what's happening is that it said, hey, you know, there's this example shader lab from Basement Studio. If I went and looked at the screen, I can even see

7:10that it was looking in my bookmarks and it was scrolling and found this um shader lab example probably down here. Yep, there we go. And Crockpot said, "Hey, I can build a demo for this in your demo repository." It ran the prompt by me. I tweaked it a little bit and then approved the build. From there it actually started a build in cursor which if we take a look at cursor cloud in chrome we see that the build is running and this is the really nice part about the integration with cursor which is that if you run things in cloud agents they sync across cursor web but then

7:45also cursor desktop. So this automation kicks off cloud agents and starts building software and I have access to that immediately. Now, I think one thing to call out which is really interesting is that there are some really unique feedback loops to this workflow, which is that if I have these automations running on these different triggers, we'll get into that. I could basically build software that's kicked off via new information, via actions that I take,

8:09via any external sort of data source. And I think that opens up the possibility for some really cool software factories and really cool um workflows. So, we'll talk about that a little bit more, but this um bot is going to kick off that build and then notify me when it's ready and actually send along some screenshots and videos. We can take a look at a previous run here where I was uh demoing some like prompt uh benchmarking um or eval uh and so Grockbot built this website, took some screenshots when I was done and opened a PR on this repository. So, that's just one example. Now, the thing

8:46to call out here is that there are three main ways you can run a bot. The first way to run a bot is by chatting with it. So, if I say, "Hey, you know, I want you to do something." The second way is by clicking in uh to this routine here and um scheduling one. Now, the easiest way to schedule a routine is just to tell the bot, "Hey, I want you to run on weekdays at 9:00 a.m." And it'll build the the trigger here. You can also build a trigger by clicking in via the guey and setting it on a schedule, setting a Slack message, setting an event from

9:15GitHub, etc. And that's also an easy way to sort of modify the name instruction set, run history, run a test run, delete the routine, or disable the routine. So that's the basics. When we create a new bot, we're really just going to get a chat interface. So let's say I create a new bot here. It's an empty chat interface. we can start to build this bot by simply explaining to Rockbot what we want to happen. Um, and it'll even kind of prompt that out of us. So, the TLDDR here and the thing I want you to remember is that if you're wondering if you can do something with Crockbot, just

9:48ask. It has a ton of different tools that it can fetch that information for, and it's going to proactively help you accomplish your goals. So, I started talking about this tech demos bot as sort of an example um and an explanation of what bots are and how they work. I think a good follow-up question is probably, well, how do I know that this

10:09thing is going to do what I ask it to? And second, what permissions does it have? Because maybe I just logged into my Twitter account here. Theoretically, that means that the bot can go to the homepage and start tweeting things, right? That might not be uh any good. How do I know that the bot won't do that? And the short answer is that the bot can do anything you give it permission to do. However, the architecture of Grockbot makes it such that we can trust the way that it's working to um sort of align with what we'd like. And we can see this if we click our profile here and go into

10:40settings. Under the bot settings, there is a allow and deny list. So you can set rules for Grockbot that it will adhere to. And basically the way that this works is that there's a reviewer agent that looks at every single action Grockbot put could take and it either says I'm going to allow Grockbot to do this thing. I'm going to deny this thing or I'm going to escalate it to the user and I'm going to ask. You can add rules here such as uh when Grockbot wants to send a tweet uh it should not allow it automatically but first ask. And that's going to prompt Grockbot to ask me

11:17before it does something. Uh, so if I was to say like, hey, send a tweet, it'll ask me before it it takes that action. Under the hood, the way this works is that actually the review list or the review agent um has these lists in context and it's proactively looking at every single message and then checking to see if um it matches one of those rules. The other thing to call out is that if I log into a website on one bot, because the underlying computer is the same across all these different

11:46bots, it will be um active on the other. So, if I open, for example, this weekly product digest bot and I open up our Chrome instance and go to x.com, I'll be logged in on this bot. It has access to Twitter the same way the other bots do. And that's just um fundamentally the way that the bots work. So, this bot, for example, is scanning Facebook Marketplace and Craigslist for espresso machines. Um, and all of my bots now have access to those two sites because I logged in with them. So, I mentioned that bots can be run in three different ways. And those are first if you send a

12:20bot a message. Second, if you define a routine or a trigger that kicks off the bot, but third, bots can also be triggered by other bots, which means that they're composable. And this has a couple different implications. The first is that you can actually create a group chat with different bots. So if I wanted the cursor product expert, the cursor product digest and then I don't know my prioritization bot in a group chat, I can do that. Now an important thing to mention, if I just say hey in this group chat, each bot is going to take turns responding. So that's one thing you should be aware of if you want to send

12:52messages to multiple bots. A better way to trigger bots sequentially or sort of on demand is to poke into one bot. So, say I jump into my writing bot and I say, "Hey, uh, look at what shipped in cursor via the um cursor product expert, then write a short tweet, uh, about um, it using my writing voice." And this is how I use that sort of composability of bots to chain actions together. So, my writing bot is essentially acting as an orchestrator and it's going to go over to the cursor product expert and say, "Hey, you know what's new?" And then it's going to bring that back into uh

13:32the um the chat here. And so we get a little message. We can see the product expert got um pinged. And then the product expert is going to start working on the query that was sent across by the writing bot. And so really that has implications for how we define workflows because we should be thinking about these bots as specialists or maybe

13:53employees that work in certain domains. And so you could have a routine, for example, that runs every weekday and routes requests through different bots to accomplish a goal. This is sort of what I meant earlier when I said there's so many possibilities that this is a totally green field product and there's a lot of different stuff to try. I can't possibly come up with all the different ideas and it presents an opportunity for you to come up with amazing ideas and share them with everyone. So we can actually see here that the cursor product expert shared a change log from um earlier today. Uh and now it's

14:24passing it over to the writing bot to draft up some responses. And that's exactly some of the workflows that I've been using this for. Okay, so to recap some of the most important features, well bots can be triggered by messages, by routines or triggers or by other bots. And that makes them composable. It means you can have group chats with bots. It means that um you can have chains of bots that work together to accomplish workflows. Bots have access to a computer under the hood that stays logged in. It has persistent sessions to the accounts you give it access to. If a bot needs you to log in, it can hand you

14:58that computer or it can request a secure input. I didn't mention that, but if a bot just needs a password, it can actually send you a secure form where you just type your password in. That means that you can take action to get bots over the hump, over uh certain difficult actions so that they can

15:14continue doing work for you. Furthermore, if you want to record a workflow, you can click record and just click through the screen to get the bot to do that. You can also provide allow or deny list rule sets to make sure that bots don't take actions you don't want them to or always take actions you do

15:30without asking you for prior approval. Setting up bots as is as easy as just sending through a prompt or a voice note. And that's often the way that I go. So from start to finish, that's kind of everything you need to know to get running with Crockbots. Oh, and we also have the plugins which allow you to connect accounts, connect services with a couple clicks and some OOTH has been rock solid for me so far. So, with that, now I want to talk about some of the workflows I've been using this tool for and give you examples um of both how I think it can be helpful and also what I

16:02think the future of software development or workflow automation looks like. Okay, so the first workflow I want to talk about, this one's actually pretty cool. I whipped up this morning on my phone. So, I have a view of it here, but we're actually going to cut over to my phone and I'll walk you through exactly what I did to build this out. I had this idea that um I've made a lot of friends on Twitter and a lot of them don't actually live in San Francisco where I live. They live all across the United States, across the world, but most of them have

16:27their location in their descriptions. That means it's public information, right? I was wondering, hey Crockbot, could you go get that information off of Twitter for each of my followers, either through the API or by physically going and looking at their profile, create a notion database, pull that information into the database, um and then hand it over to me. Now, typically this is something we might use AI to write a script for to accomplish that goal, right? This is typically like a personal scripting or software approach. Now, I just had an agent do it and it's kind of better because it's completely disposable. Um, it took me maybe 5 or 10

17:03minutes and now I have access to this agent forever. I can tweak it. I can update it and I can approve it. So, you were looking at the mobile view. Um, I'll pop open Notion right here. And I follow roughly eight or nine00 people. Now I have um a list of all of them here with profile images, which is kind of nice. Uh, their descriptions and where they're located, the country, the um a geo confidence here and then like kind of a fuzzy location. So, you know, say uh, you know, I'm a San Francisco guy and I'm going to New York. I could search here New York um, and I could get

17:37all of my followers that reside in New York. and then I could ping people and be like, "Hey, you want to grab a coffee? I'm the annoying coffee guy out here asking people to grab a coffee." This, I mean, quite literally, I spun up in 15 minutes. Uh, actually, I think closer to 10 minutes um with my phone this morning. And that means it's less about uh what you can build and more about what you can think or what you can sort of dream up, what your imagination

18:02contains. Basically, take anything that you could do manually. Hey, I'm going to go like go through all my LinkedIn follow or LinkedIn connections and do this thing. And just think what if I gave this to an agent and just added to it manually. And I think that's really amazing and really cool. So that's one workflow example that I did on my own. Okay. Second, and I think this one is actually a little cooler. In front of me is a piece of software. I didn't build this with Grockpot. I actually vibecoded this mostly with Fable 5, but I spent a lot of time building out this uh personal

18:30training coach. So it allows me to period periodize my training and use some principles of uh you know scientific um hypertrophy and strength training which is something I'm very passionate about to build a program and then from that program I can go through and log my exercises I can get a projection for the next week and that's great but what I realized is that you know this is a piece of software and if you build software then you have to maintain it and a lot of times this breaks or there'll be stuff that's incorrect and the way I got to this app was by like weeks and honestly months of

19:01building building on top of what was relatively simple progression logic and a shared data source and that brings us a little bit back to like what software is. So if you think about software, it's really like three things. It is logic. I, you know, have some business logic or some computations or things I want to happen with some data, which is the second piece that I'm storing persistently either in a database or a JSON file or, you know, object storage, but really data can be anything. It can be a CSV, it can be something in Google Drive, it can be in GitHub, it can be on

19:33your computer. So I'm storing some data and I'm performing logic on that and then I'm interacting with that software through an interface. So you could think about it as well like there is a server that is performing those computations. There's a persistent data store and there's a UI client server database right that's that's software kind of in a nutshell most full stack applications. Well, I thought, what if I brought that software into chat? And I think what I've started to realize is that agents are software. And so, what I did with this um Arnold bot here is that under the hood, this is my strength and hypertrophy programming

20:11coach. Now, it's possible many of you have just used chatbt or grock or codeex or whatever claude for this, but the reason this is different is because I gave this um agent access to all of the logic that was in my app. So, I took my app, I broke it down into a set of MCP servers into a set of skills and plugins, and then I prompted Grockbot uh to build all of this. And I actually did all of this from inside Grock. So, if we go back up here, you can actually see I'm adding periodization plugins and skills incursor. There was an error there, but took one more try. And I

20:48could open that in cursor as well and take a look at the actual PR and see what I built and then committed to Grockbot. And then I'm just using the same interface to chat with Grockbot. And it's sending through all of my different workouts. Under the hood, it's logging these in git as commits to, I believe, a JSON file. But what I have is a piece of software with logic and a computer. The agent is the computer and the interface is chat. So, while this might not look as beautiful or as fullfeatured as what I had before, this is much more easily maintainable. And if there's an error,

21:26the bot is actually going to be able to diagnose that, write code, commit to the GitHub repository, and publish that. And that's a much faster feedback loop than my training app, right? Because in my training app, I'd be basically have to fire off cursor cloud agents. I'd have to go home. I'd have to, you know, maybe look at things, approve, deny, merge, set up this like preview, deploy, feedback loop, which is maybe a little bit tricky depending on the technologies that you that you use. And so this is a fundamentally simpler approach to software whereby the feedback loop of iterating on that software is in the

21:58chat. Another thing I kind of wanted to build AI chat into my training app. I wanted to be able to chat with it as if it was a coach. With Crockpot, I already have that out of the box. So, depending on what the software is that you're trying to build, it might just be an agent. And we're very soon going to ship multiplayer for tools like Rockbot that's going to allow you to basically share this software with all your teammates. So, this is the personal software approach and something to keep in mind. If you can break your app down into a client, a server, and a database,

22:33consider attaching some sort of data source to Grock, either by connecting it to Google Drive, giving it access to a git repo, or just giving it some files. It has a computer and then using the agent itself as the computer, either by giving it skills, and you can also just define skills for Grockbot. Something we didn't talk about, um, you can attach files, teach a task, but then Grockbot also has access to skills if you define them. um MCP servers tools plugins um and then using chat as the input and output that may be a replacement for a lot of the personal software or personal scripts you've been building. Okay,

23:08another way I've been using Grockbot is to write code. Now, we already talked about the sort of Xbookmarks example where I'm building demos off of my Twitter bookmarks and I'm dispatching agents that way. Now I want to talk about a bit more of an advanced concept which is doing actual software engineering from inside Grockbot and I think Lauren part of the cursor team really amazing engineer um stated this best. I sat down with her at a conversation and she described Grockbot to me as the perfect outer loop. And so

23:38I actually call this agent outer loop. Basically, what we want is some outer loop where we're gathering context, understanding problems, pulling in context from these different systems before we kick it off to a cloud agent that acts as our inner loop and actually builds the software. Why do we do this? Well, we do it to protect the context of the inner loop, the cloud agent, um, and to give us a place to explore. Why is this important? Well, LLM's work on next token prediction, right? That means that everything that's in context is going to influence the outcomes. So if you're in an agent and you're like, "Oh, read

24:10these files. What should I build? Should I build this? Help me think through these problems." And then from that point, you start delegating fixes. That agent is actually going to have a bunch of dirty context and a bunch of information that might not be relevant to what you're trying to build. So with this outer loop in loop uh architecture, we can protect the context of that inner loop and focus on collecting information with essentially an agent that's specialized for that. So this agent I have here uh and if we look at the description, this is an expert on the cursor codebase, documentation, and marketing repos. So I gave it access to

24:45all of those things. And its goal is to help me look through sources of context and come up with perfect prompts to then delegate to cloud agents with the skills that I use to build at cursor. One of those is Lauren's um Pstack uh plugin, which is amazing. So an example prompt that I might send through here is do research on user credit grants. this is a problem I've been thinking about. Look across notion and Slack and see how our company engages with granting user credits. Then come up with a plan to improve the system and kick off a prompt to a cloud agent. So what this bot is

25:15going to do is going to go through all of those sources, start to understand the problem. Ideally, it would come back to me. We'd chat through it and then the same thing we did um down here with our tech demos, it's going to fire off uh a cursor cloud agent to go out and actually build the thing. Again, the goal of this and why we might do it is to priorit prioritize and protect the context um of that interloop coding agent and keep it more specialized. So, a bit of like meta discussion here, a bit of feature of building, but I've been using this flow to basically kick off the the bots that

25:52I have working in cursor. So, a lot of my cursor um bots end up started from Crockbot. We're not replacing the code coding harness. We're just augmenting it and staging for that coding harness. Another really great use case, I won't dive too deep on this one, is just a cheap of staff or prioritization bot. So, I use this kind of for a catch-all of like, hey, here are some really good links on Devril. Can you put these in my

26:18notion? We'll just write to my database. It also has access to my uh Gmail and my calendar, so I can have it draft responses to emails, which has helped me automate a few of those things. And then I do a weekly review and a Monday weekly kickoff as well and it prompts me to do those because otherwise I just don't actually do it, you know, so I have a

26:37bot keeping me accountable essentially. So that's another really great use there. So we talked a little bit about coding, we talked about personal software, we talked about workflows, now I want to talk about information retrieval, which is maybe one of the simpler ways to use Grockbot, but honestly insanely helpful. And this is how I onboarded to cursor. Um, if you want to see a write up of this, I have a Twitter article that I shared out um, not too long ago. We could go to x.com

27:06and the article is chat is all you need. And I kind of talk about, you know, um, lower on down here. Uh, we have information retrieval. I talk about how difficult it can be uh, to access information across different resources. You might have felt this if your company has a notion, a Slack, a different wiki, uh you know, you're responding to emails, it can be really hard to find stuff. And so my onboarding at Cursor was largely driven by Grockbot. And I used um actually some combination of this Cursor product expert bot um number one to answer questions about the product, but then also to help me answer

27:46how the company works. So, you can even see when I was doing research for this for this um video, I said, "Hey, how do Grockbot group chats work?" It has access to the codebase. It could just go and find that for me and tell me that uh all the agents that you add join the room. There's an orchestrator, but it's not one of them. And then the host runs

28:03a round robin after you send a message. So, it wakes each member in turn. That's how I knew what to tell you all, right? And I've been using this to learn about the products, but also to search across Slack and Notion. My onboarding really took like a couple of days. Uh, and it's still ongoing, right? But I feel like I can answer all these questions because all of that information is accessible and saved off somewhere. All right, so I've been talking a lot, but the main things to remember here are that one, Grock is an agent that's running in the cloud. Uh, Grockbot is an agent with a computer,

28:36and it can kind of do anything you can do. So you really have to kind of expand what's possible here and think about writing code, running code, running software, doing workflows, controlling a computer the same way you can. And if you delegate appropriately, you're going to be able to accomplish quite a bit with this tool. You can control what Grockbot is able to do or not do. And you can access Grockbot through your phone or your desktop computer. It runs in the cloud, so you can close your laptop. You don't have to worry about it shutting off. It's a persistent Linux VM uh under the hood on a server that's

29:08running in the cloud. The main difference between Grockbot and many of these other personal agents you see out there are that there's zero setup and you don't have to do any server management. You don't have to do any configuration. You can just jump in and start building and that's really impactful. So, I highly recommend giving it a shot. Check it out. Let me know what you think. Let me know if this video helped you. But until next time,

29:30I'm Matt with SpaceX. Peace.