← Back to videos
This is f*cking insane. Elon Musk keeps telling everyone to try Grok Bot. Matt Palmer made the guide

David Keln
@0xdav1d_k
November 18, 2026· 30:20
This is f*cking insane. Elon Musk keeps telling everyone to try Grok Bot. Matt Palmer made the guide: "Grok Bot can do anything you can do on a computer One bot chains into the next, and the whole thing runs while you sleep" In 30 minutes, he builds 4 real workflows live: X followers into Notion, personal software, coding with Cursor and search What Is a Bot → Anatomy → Multi-Bot Chains → 4 Real Workflows 1.4 million people watched it. Most of them never built a single workflow Worth more than any $1,500 agent course Bookmark and watch it today Then read the 6 Grok bots, 200+ agents setup ↓
Transcript
- Hey, Grokbot
- could you go get that information
- 0:03
- off of Twitter for each of my followers
- 0:04
- either through the API
- 0:06
- or by physically going and looking at their profile
- 0:09
- create a Notion database
- 0:10
- pull that information into the database
- 0:13
- and then hand it over to me
- 0:14
- Now
- 0:14
- typically this is something we might use AI
- 0:16
- to write a script for to accomplish that goal
- 0:19
- right? This is typically like a personal
- 0:22
- scripting or software approach
- 0:24
- Now I just had an agent do it
- 0:26
- and it's kind of better because it's completely disposable
- 0:31
- it took me maybe five or ten minutes
- 0:33
- and now I have access to this agent
- 0:34
- forever. I can tweak it
- 0:36
- I can update it
- 0:37
- and I can approve it
- 0:37
- Today we're going to talk all about Grokbot
- 0:39
- Specifically, we'll talk about what Grokbot is and how it works
- 0:43
- and then we're going to move into how
- 0:44
- you can get started and how you can
- 0:46
- set up the tool
- 0:47
- to start automating your own workflows
- 0:49
- And finally, we're going to dig a bit deeper
- 0:50
- and I'll show you some of my favorite
- 0:51
- workflows and uses for this tool
- 0:53
- because really, the possibilities are endless
- 0:57
- as cliche as that might sound
- 0:59
- and we're all Still kind of figuring out
- 1:00
- what this tool is capable of
- 1:02
- So
- 1:03
- what is Grokbot? It might be easier to
- 1:05
- show you rather than explain
- 1:06
- I have
- 1:07
- a bot open in front of me here
- 1:09
- Under the hood
- 1:09
- Grokbot is just an agent with a computer
- 1:12
- So an agent, as you're familiar with
- 1:15
- is just a model that has access to
- 1:16
- some tools and can do stuff for us
- 1:19
- And this agent has access to a computer
- 1:21
- So if I click this window
- 1:23
- we can actually see the computer
- 1:24
- More than that
- 1:25
- we can interact with it like a remote
- 1:26
- desktop
- 1:27
- So behind the scenes
- 1:29
- this is running on a virtual machine in
- 1:31
- the cloud
- 1:32
- somewhere. So none of this is happening on
- 1:33
- my laptop
- 1:34
- And I have a terminal in front of
- 1:35
- me
- 1:36
- I could run a command
- 1:39
- I have a file system
- 1:40
- and I have a X here
- 1:43
- Now the cool thing is that Grokbot has
- 1:45
- access to
- 1:46
- all of these things the same way
- 1:49
- we do
- 1:50
- So more than just an agent with a computer
- 1:52
- it's an agent with a computer
- 1:54
- with really good computer use
- 1:56
- And we can actually
- 1:58
- teach GrokBot how to do certain things
- 2:00
- on this computer by recording and teaching a task to GrokBot
- 2:05
- or by prompting and instructing it
- 2:08
- And so today
- 2:09
- I'm going to go over
- 2:10
- some workflows and some things that I've found
- 2:12
- GrokBot to be really good at
- 2:14
- But just to give you a bit of a taste there
- 2:17
- that might include things like doing grocery shopping
- 2:20
- For example
- 2:21
- I asked GrokBot to
- 2:22
- look at
- 2:23
- Costco
- 2:23
- and look at Amazon
- 2:24
- and add certain things to my cart here
- 2:26
- and then compare prices
- 2:28
- compare delivery charges, compare times, et cetera
- 2:33
- And it's really good at stuff like that
- 2:34
- We can control the computer just like GrokBot
- 2:37
- can. So
- 2:38
- when it needs us to step in to
- 2:39
- complete two-factor authentication or log into a website
- 2:42
- it can hand that computer off to us
- 2:43
- and allow us to complete that action
- 2:46
- So, recap. What is GrokBot?
- 2:48
- It is an agent with a computer
- 2:49
- that runs in the cloud
- 2:51
- so nothing's running on your machine
- 2:52
- and that gives us access
- 2:53
- via mobile
- 2:54
- and via desktop
- 2:55
- that has really good computer use
- 2:57
- And that's really it
- 2:58
- that's the foundation. So while the inputs and outputs are chat
- 3:03
- we're always typing text through GrokBot
- 3:05
- those are pretty simple
- 3:07
- Under the hood
- 3:07
- there's this whole like operating system
- 3:10
- You can think of like as a
- 3:11
- server
- 3:12
- or a really powerful tool that allows us
- 3:14
- to get stuff done
- 3:15
- And that's the foundation
- 3:17
- for the rest of this video
- 3:19
- So just to drive this home
- 3:20
- I can come over to my marketplace bot
- 3:22
- here and ask it to do a search
- 3:24
- for a variegated monstera albo
- 3:27
- and what we'll see in the computer use on the right
- 3:31
- is that the bot
- 3:32
- because I've already authorized and logged into Facebook
- 3:35
- is going to be able to open up that machine
- 3:39
- go to a new tab
- 3:40
- or maybe use the same one
- 3:41
- and do a new search
- 3:42
- It's going to take all the actions that
- 3:43
- we would take
- 3:44
- So really, it's more than a personal assistant
- 3:47
- it's more than some of the other AI
- 3:48
- tools you're familiar with
- 3:50
- It's an agent that can do everything that
- 3:52
- you can do on a computer
- 3:54
- So we can see GrokBot performing
- 3:56
- searches for a few different plants
- 3:58
- and only finding a few of them
- 4:00
- on Craigslist. So what can GrokBot do?
- 4:03
- I want to give you a few use
- 4:04
- cases before we dig into
- 4:05
- how to use GrokBot and how to get things set up
- 4:08
- so you kind of understand why you might
- 4:10
- adopt a tool like this if you're still
- 4:11
- curious
- 4:12
- I've been using Grokbot for more main
- 4:14
- four main things, which is first, writing software, second
- 4:18
- building personal software as agents
- 4:21
- This is a little bit different
- 4:22
- I'm actually moving the software into chat
- 4:24
- We'll talk about what that means
- 4:26
- later on in the video
- 4:28
- for information and knowledge retrieval
- 4:30
- so looking across the cursor in SpaceX
- 4:33
- Slacks, Notion, GitHub, it's been really helpful
- 4:36
- It's actually how I onboarded to cursor
- 4:38
- We'll use that as an example
- 4:40
- And fourth, for automating workflows
- 4:42
- So very similar to what I just showed
- 4:43
- you with this grocery example
- 4:45
- you can think of any workflow
- 4:47
- any sort of knowledge work you might have to do
- 4:49
- whether that's
- 4:50
- going on LinkedIn and reaching out to people or finding information
- 4:53
- if you're a salesperson
- 4:54
- bring that into your CRM
- 4:56
- reaching out to these clients
- 4:57
- drafting emails, et cetera
- 4:59
- A lot of our go-to-market team actually does
- 5:01
- just that with Grokbot
- 5:03
- Similarly, for our most recent launch
- 5:05
- I actually used Grokbot to
- 5:07
- coordinate across all the different feedback we're receiving
- 5:10
- for our launch in Slack
- 5:12
- in Notion
- 5:13
- and then to sync all the content between Notion to PRs
- 5:17
- to incorporate Notion comments
- 5:19
- PR comments, Slack feedback
- 5:21
- and then to kick off cursor cloud agents
- 5:23
- to
- 5:23
- make sure that all the feedback was integrated
- 5:26
- And it saved me a tremendous amount of time
- 5:28
- and it was a really great use case
- 5:29
- So we'll dig into those four things
- 5:31
- but first I want to show you how
- 5:33
- to set
- 5:34
- things up
- 5:34
- So what are these agents and how do they work?
- 5:37
- Well, let's take a look at an example
- 5:39
- and this is one of my favorite
- 5:40
- my tech demos example
- 5:42
- And if I open up the settings here inside the bot
- 5:45
- we can see there's a name
- 5:47
- a title, and a description
- 5:49
- The
- 5:50
- name is technically
- 5:52
- optional and doesn't actually influence the behavior of
- 5:54
- the bot
- 5:55
- The title is a description of what the bot does
- 5:57
- and then the
- 5:58
- actual description is a longer form explanation
- 6:02
- So this bot looks at my Twitter bookmarks
- 6:06
- I call it like colloquially the DevRel bot
- 6:09
- and each day
- 6:10
- comes up with something that's cool
- 6:12
- out of the Bookmarks
- 6:13
- From there, it drafts up a prompt in my style
- 6:16
- sends it over to me
- 6:16
- and then I can approve or deny that
- 6:18
- prompt. Behind the scenes
- 6:20
- it's going to take that and kick it
- 6:21
- off to a cursor cloud agent
- 6:22
- through the cursor integration
- 6:24
- And that brings me to another really powerful part of Grokbot
- 6:27
- which is that it supports all of the integrations
- 6:30
- MCP servers, and plugins that cursor does
- 6:33
- as well as skills
- 6:34
- et cetera
- 6:35
- So we can add things like Gmail
- 6:38
- Google Calendar, Google Drive
- 6:40
- some of the more complex ones and internal
- 6:43
- things that our team are using here
- 6:44
- as well as other sasses like Amplitude for data analytics
- 6:48
- Twitter for accessing more structured data
- 6:51
- 1Password, et cetera
- 6:53
- And the nice thing
- 6:54
- is that we can even add multiple accounts
- 6:56
- So I have multiple Gmail accounts there
- 6:58
- If I go to actually Notion
- 7:01
- which I think is the better example
- 7:03
- I can have my SpaceX
- 7:04
- cursor, and personal accounts
- 7:06
- We can retrieve information across all those sources
- 7:09
- So this has cursor as A connection
- 7:13
- And what's happening is that it said
- 7:15
- hey, you know, there's this example
- 7:17
- shader lab from basement studio
- 7:19
- If I went and looked at the screen
- 7:21
- I can even see that it was looking
- 7:23
- in my bookmarks and it was scrolling and found this
- 7:27
- shader lab example, probably down here
- 7:29
- Yep, there we go
- 7:31
- And Grokbot said, hey
- 7:34
- I can build a demo for this in
- 7:35
- your demo repository
- 7:37
- It ran the prompt by me
- 7:38
- I tweaked it a little bit
- 7:39
- and then approved the build
- 7:40
- From there, it actually started a build in cursor
- 7:44
- which if we take a look at cursor cloud in Chrome
- 7:48
- we see that the build is running
- 7:49
- And this is the really nice part about
- 7:51
- the integration with cursor
- 7:53
- which is that if you run things in cloud agents
- 7:55
- they sync across the cursor web
- 7:57
- but then also cursor desktop
- 7:59
- So this automation kicks off cloud agents and starts building software
- 8:02
- and I have access to that immediately
- 8:04
- Now, I think one thing to call out
- 8:06
- which is really interesting
- 8:07
- is that there are some really unique feedback
- 8:09
- loops to this workflow
- 8:10
- which is that if I have these automations
- 8:13
- running on these different triggers
- 8:15
- we'll get into that
- 8:17
- I could basically build software that's kicked off via new information
- 8:21
- via actions that I take
- 8:22
- via any external sort of data source
- 8:26
- And I think that opens up the possibility
- 8:28
- for some really cool software factories and really
- 8:30
- cool
- 8:31
- workflows
- 8:32
- So we'll talk about that a little bit
- 8:33
- more
- 8:33
- But this
- 8:36
- bot
- 8:36
- is going to kick off that build and
- 8:38
- then notify me when it's ready
- 8:40
- and actually send along some screenshots and videos
- 8:42
- We can take a look at a previous
- 8:43
- run here
- 8:44
- where I was demoing some like prompt
- 8:47
- benchmarking or evals
- 8:50
- And so Grokbot built this website
- 8:52
- took some screenshots when it was done
- 8:54
- and opened a PR on this repository
- 8:56
- So that's just one example
- 8:59
- Now
- 9:00
- The thing to call out here
- 9:01
- is that
- 9:02
- there are three main ways you can run
- 9:04
- a bot
- 9:04
- The first way to run a bot is
- 9:06
- by chatting with it. So if I say
- 9:07
- hey, you know
- 9:07
- I want you to do something
- 9:08
- The second way is by clicking in
- 9:11
- to this routine here and
- 9:14
- scheduling one. Now
- 9:15
- the easiest way to schedule a routine is
- 9:17
- just to tell the bot
- 9:18
- hey, I want you to run on weekdays at 9 AM
- 9:20
- and it'll build the
- 9:21
- the trigger here
- 9:23
- You can also build a trigger by clicking
- 9:25
- in via the GUI and setting it on a schedule
- 9:27
- setting a Slack message
- 9:29
- setting an event from GitHub
- 9:31
- et cetera
- 9:32
- And that's also an easy way to sort
- 9:33
- of modify the name
- 9:35
- instruction set, run history, run a test run
- 9:38
- delete the routine, or disable the routine
- 9:41
- So
- 9:43
- That's the basics. When we create a new bot
- 9:45
- we're really just going to get a chat
- 9:47
- interface
- 9:47
- So let's say
- 9:48
- I create a new bot here
- 9:49
- It's an empty chat interface
- 9:50
- We can start to build this bot by
- 9:52
- simply explaining to RockBot what we want to happen
- 9:56
- and it'll even kind of prompt that out
- 9:58
- of us. So
- 9:59
- the
- 9:59
- TLDR here and the thing I want you
- 10:01
- to remember is that
- 10:02
- if you're wondering if you could do something with RockBot
- 10:04
- just ask. It has a ton of different
- 10:06
- tools that it can fetch that information for
- 10:09
- and it's going to proactively
- 10:11
- help you accomplish your goals
- 10:13
- So
- 10:13
- I started talking about this tech demos bot
- 10:16
- as sort of an example
- 10:18
- and an explanation of what bots are and
- 10:20
- how they work
- 10:22
- I think a good follow-up question is probably
- 10:24
- well
- 10:25
- how do I know that this thing is
- 10:26
- going to do what I ask it to?
- 10:28
- And second, what permissions does it have?
- 10:30
- Because maybe I just logged into my Twitter account here
- 10:33
- Theoretically
- 10:34
- that means that the bot can go to
- 10:35
- the homepage and start tweeting things
- 10:37
- right?
- 10:37
- That might not be any good
- 10:40
- How do I know
- 10:41
- that the bot won't do that?
- 10:43
- And the short answer is that the bot
- 10:44
- can do anything you give it permission to do
- 10:47
- However
- 10:47
- the architecture of GrokBot makes it such that
- 10:50
- we can trust the way that it's working to
- 10:53
- sort of align with what we like
- 10:55
- And we can see this if we click
- 10:56
- our profile here and go into settings
- 10:59
- Under the bot settings
- 11:00
- there is a allow and deny list
- 11:03
- So
- 11:03
- you can set rules
- 11:05
- for GrokBot that it will adhere to
- 11:08
- And
- 11:08
- basically the way that this works is that
- 11:10
- there's a reviewer agent that looks at every
- 11:12
- single action GrokBot could take
- 11:15
- and it either says
- 11:16
- I'm going to allow GrokBot to do this thing
- 11:18
- I'm going to deny this thing
- 11:20
- or I'm going to escalate it to the user
- 11:22
- and I'm going to ask
- 11:23
- You can add rules here
- 11:25
- such as, when GrokBot wants to send a tweet
- 11:29
- it should not allow it automatically
- 11:31
- but first ask
- 11:33
- And that's going to prompt Grokbot to
- 11:35
- ask me before it does something
- 11:38
- so if I was to say like
- 11:39
- hey, send a tweet
- 11:40
- it'll ask me before it takes that action
- 11:42
- Under the hood
- 11:43
- the way this works is that actually the review list
- 11:46
- or the review agent
- 11:47
- has these lists in context
- 11:50
- and it's
- 11:51
- proactively looking at every single message
- 11:53
- and then checking to see if
- 11:56
- it matches one of those rules
- 11:58
- The other thing to call out is that
- 11:59
- if I log into a website on one bot
- 12:01
- because the underlying computer is the same across
- 12:04
- all these different bots
- 12:06
- it will be, active on the other
- 12:08
- So if I open
- 12:09
- for example, this weekly product digest bot
- 12:11
- and I open up our Chrome instance and go to x.com
- 12:14
- I'll be logged in on this bot
- 12:16
- It has access to Twitter
- 12:18
- the same way the other bots do
- 12:20
- And that's just, Fundamentally
- 12:22
- the way that the bots work
- 12:24
- So this bot, for example
- 12:26
- is scanning Facebook Marketplace and Craigslist for espresso machines
- 12:31
- and all of my bots now have access
- 12:33
- to those two sites because
- 12:34
- I logged in with them
- 12:35
- So I mentioned
- 12:36
- that bots can be run in three different ways
- 12:38
- and those are first
- 12:39
- if you send a bot a message
- 12:41
- second
- 12:41
- if you define a routine or a trigger
- 12:43
- that kicks off the bot
- 12:44
- but third, bots can also be triggered by other bots
- 12:47
- which means that they're composable
- 12:49
- And this has a couple different implications
- 12:52
- The first is that you can actually create
- 12:54
- a group chat with different bots
- 12:55
- So if I wanted the cursor product expert
- 12:57
- the cursor product digest
- 13:00
- and then, I don't know
- 13:01
- my prioritization bot in a group chat
- 13:03
- I can do that. Now
- 13:04
- an important thing to mention
- 13:05
- if I just say hey in this group chat
- 13:07
- each bot is going to take turns responding
- 13:10
- So that's
- 13:10
- one thing
- 13:11
- you should be aware of if you want
- 13:12
- to send messages to multiple bots
- 13:15
- A better way to trigger bots sequentially
- 13:18
- or sort of on demand
- 13:20
- is to poke into one bot
- 13:21
- So say I jump into my writing bot and I say
- 13:24
- hey, look at what shipped in Cursor via the
- 13:30
- Cursor product expert, then write a short tweet
- 13:35
- about, It using my writing voice
- 13:39
- And this is how I use that sort
- 13:41
- of composability of bots to chain actions together
- 13:44
- So my writing bot is essentially acting as an orchestrator
- 13:47
- and it's going to go over to the
- 13:48
- cursor product expert and say
- 13:49
- hey, you know, what's new?
- 13:51
- And then it's going to bring that back into
- 13:54
- the, the chat here. And so we get a little message
- 13:58
- we can see the product expert got
- 14:00
- pinged
- 14:01
- and then
- 14:02
- the product expert
- 14:03
- is going to start working on the query
- 14:06
- that was sent across by the writing bot
- 14:08
- And so really that has implications for how we define workflows
- 14:12
- because we should be thinking about these bots
- 14:14
- as specialists or maybe employees that work in
- 14:17
- certain domains
- 14:18
- And so you can have a routine
- 14:20
- for example
- 14:20
- that runs every weekday
- 14:22
- and routes requests through different bots to accomplish
- 14:25
- a goal
- 14:26
- This is sort of what I meant earlier
- 14:27
- when I said there's so many possibilities that
- 14:29
- this is a
- 14:30
- totally greenfield product and there's a lot of
- 14:32
- different stuff to try
- 14:34
- I can't possibly come up with all the
- 14:35
- different ideas and it presents an opportunity for
- 14:37
- you to come up with some amazing
- 14:39
- ideas and share them with everyone
- 14:40
- So we can actually see here that the
- 14:41
- cursor product expert shared
- 14:43
- a change log from earlier today
- 14:46
- and now it's passing it over to
- 14:49
- the writing bot to draft up some responses
- 14:52
- And that's exactly
- 14:54
- some of the workflows that I've been using this for
- 14:56
- Okay, so to recap, some of the most important features
- 14:59
- Well, bots can be triggered by messages
- 15:01
- by routines or triggers
- 15:03
- or by other bots
- 15:04
- and that makes them composable
- 15:06
- It means you can have group chats with
- 15:07
- bots
- 15:08
- It means that
- 15:09
- you can have chains of bots that work
- 15:11
- together to accomplish workflows
- 15:13
- Bots have access to a computer under the
- 15:15
- hood that stays logged in
- 15:16
- It has persistent sessions to the accounts you
- 15:18
- give it access to
- 15:20
- If a bot needs you to log in
- 15:22
- it can hand you that computer
- 15:23
- or it can request a secure input
- 15:25
- I didn't mention that
- 15:26
- but if a bot just needs a password
- 15:28
- it can actually send you a secure form
- 15:30
- where you just type your password in
- 15:32
- That means that you can take action to
- 15:34
- get bots over the hump
- 15:36
- over, certain difficult actions
- 15:38
- so that they can continue doing work for you
- 15:40
- Furthermore, if you want to record a workflow
- 15:43
- you can click record and just click through
- 15:45
- the screen
- 15:45
- to get the bot to do that
- 15:47
- You can also provide allow or deny list
- 15:49
- rule sets to make sure that bots don't
- 15:51
- take actions you don't want them to
- 15:53
- or always take actions you do
- 15:55
- without asking you for prior approval
- 15:58
- Setting up bots
- 15:59
- is as easy as just sending through a
- 16:00
- prompt or a voice note
- 16:02
- and that's often the way that I'd go
- 16:04
- So From start to finish
- 16:06
- that's kind of everything
- 16:07
- you need to know to get running with GrokBots
- 16:10
- Oh, and we also have the plugins
- 16:11
- which allow you to connect accounts
- 16:13
- connect services with a couple of clicks
- 16:15
- and some OAuth
- 16:16
- has been rock solid for me so far
- 16:19
- So with that
- 16:20
- now I want to talk about some of
- 16:21
- the workflows I've been using this tool for
- 16:24
- and give you examples
- 16:26
- of both how I think it can be
- 16:27
- helpful and also
- 16:28
- what I think the future
- 16:29
- of software development or workflow automation looks like
- 16:32
- Okay, so the first workflow I want to talk about
- 16:35
- this one's actually pretty cool
- 16:36
- I whipped up this morning on my phone
- 16:37
- So I have a view of it here
- 16:39
- but we're actually going to cut over to
- 16:40
- my phone and I'll walk you through exactly
- 16:41
- what I did to build this out
- 16:43
- I had this idea that
- 16:45
- I've made a lot of friends on Twitter
- 16:47
- and a lot of them don't actually live
- 16:48
- in San Francisco where I live
- 16:50
- They live all across the United States
- 16:51
- across the world
- 16:53
- but most of them have their location
- 16:54
- in their descriptions. That means it's public information
- 16:57
- right?
- 16:57
- I was wondering, hey, GrokBot
- 16:59
- could you go get that information
- 17:01
- off of Twitter for each of my followers
- 17:03
- either through the API
- 17:04
- or By physically going and looking at their profile
- 17:07
- create a Notion database
- 17:09
- pull that information into the database
- 17:11
- and then hand it over to me
- 17:12
- Now
- 17:13
- typically this is something we might use AI
- 17:15
- to write a script for to accomplish that goal
- 17:18
- right? This is typically like a personal
- 17:20
- scripting or software approach
- 17:23
- Now I just had an agent do it
- 17:24
- and it's kind of
- 17:26
- better because it's completely disposable
- 17:30
- It took me maybe five or ten minutes
- 17:31
- and now I have access to this agent
- 17:33
- forever. I can tweak it
- 17:34
- I can update it
- 17:35
- and I can approve it
- 17:36
- So you were looking at the mobile view
- 17:38
- I'll pop open Notion right here
- 17:40
- and
- 17:41
- I follow roughly eight or nine hundred people
- 17:43
- Now I have
- 17:45
- a list of all of them here with profile images
- 17:47
- which is kind of nice
- 17:49
- Their descriptions and where they're located
- 17:52
- the country, the, a geoconfidence here
- 17:56
- and then like kind of a fuzzy location
- 17:58
- So, you know, say, you know
- 18:00
- I'm a San Francisco guy and I'm going to New York
- 18:03
- I could search here New York
- 18:05
- and I could get all of my followers
- 18:07
- that reside in New York
- 18:08
- and then I could ping people and be like
- 18:10
- hey, you want to grab a coffee?
- 18:12
- I'm the annoying coffee guy out here
- 18:14
- asking people to grab a coffee
- 18:16
- This, I mean, quite literally, I spun up in 15 minutes
- 18:19
- actually I think closer to 10 minutes
- 18:22
- with my phone this morning
- 18:23
- And that means it's less about
- 18:26
- what you can build and more about what
- 18:28
- you can think or what you can sort of dream up
- 18:30
- what your imagination contains
- 18:33
- Basically take anything that you could do manually
- 18:36
- Hey, I'm going to go like
- 18:37
- go through all my LinkedIn follower
- 18:38
- or LinkedIn connections and do this thing
- 18:41
- And just think
- 18:41
- what if I gave this to an agent
- 18:42
- and just had it do it manually?
- 18:44
- And I think that's really amazing
- 18:46
- and really cool
- 18:47
- So that's one workflow example that I did on my
- 18:49
- okay
- 18:50
- second. And I think this one is actually
- 18:51
- a little cooler
- 18:52
- In front of me is a piece of
- 18:53
- software. I didn't build this with Grokbot
- 18:55
- I actually vibe coded this
- 18:56
- mostly with Fable 5
- 18:57
- But I spent a lot of time building
- 18:59
- out this personal training coach
- 19:01
- So it allows me to
- 19:02
- periodize my training and use some principles of
- 19:06
- you know, scientific hypertrophy and strength training
- 19:09
- which is something I'm very passionate about
- 19:10
- to build a program
- 19:12
- And then from that program
- 19:13
- I can go through and log my exercises
- 19:15
- I can get a projection for the next
- 19:17
- week. And that's great
- 19:18
- But what I realized is that
- 19:20
- you know, this is a piece of software
- 19:22
- And if you build software
- 19:23
- then you have to maintain it
- 19:25
- And a lot of times this breaks
- 19:26
- or there'll be stuff that's incorrect
- 19:28
- And the way I got to this app
- 19:29
- was by like weeks
- 19:31
- and honestly months of building
- 19:33
- Building On top of
- 19:34
- what was relatively simple progression logic
- 19:37
- and a shared data source
- 19:39
- And that brings us a little bit back
- 19:40
- to like what software is
- 19:42
- So if you think about software
- 19:43
- it's really like three things
- 19:44
- It is logic
- 19:46
- I, you know
- 19:46
- have some business logic or some
- 19:48
- computations or things I want to happen with some data
- 19:52
- which is the second piece that I'm storing persistently
- 19:54
- either in a database or a JSON file or
- 19:58
- you know, object storage
- 19:59
- but really data can be anything
- 20:00
- It can be a CSV
- 20:01
- it can be something in Google Drive
- 20:04
- it can be in GitHub
- 20:05
- it can be on your computer
- 20:06
- So I'm storing some data and I'm performing logic on that
- 20:09
- and then I'm interacting with that software through
- 20:11
- an interface
- 20:13
- So you could think about it as well
- 20:15
- like there is a server that is performing those computations
- 20:18
- there's a persistent data store
- 20:20
- and there's a UI
- 20:21
- client server database, right? That's
- 20:22
- that's software kind of in a nutshell
- 20:24
- most, most full stack applications
- 20:26
- Well, I thought
- 20:27
- what if I brought that software into chat?
- 20:30
- And I think what I've started to realize
- 20:31
- is that agents are software
- 20:34
- And so what I did with this
- 20:37
- Arnold bot here is that under the hood
- 20:40
- this is my strength and hypertrophy programming coach
- 20:44
- Now it's possible many of you have just
- 20:46
- used ChatGPT or Grok or Codex or whatever
- 20:49
- Claude for this
- 20:51
- But the reason this is different is because
- 20:53
- I gave this
- 20:55
- agent access to all of the logic that
- 20:58
- was in my app
- 20:59
- So I took my app
- 21:00
- I broke it down into a set of MCP servers
- 21:03
- into a set of skills and plugins
- 21:06
- and then I prompted Grokbot to
- 21:09
- build all of this. And I actually did
- 21:10
- all of this from inside Grok
- 21:12
- So if we go back up here
- 21:14
- you can actually see
- 21:15
- I'm adding periodization plugins and skills in cursor
- 21:19
- There was an error there
- 21:19
- but it took one more try
- 21:21
- And I could open that in cursor as
- 21:23
- well and take a look at the actual
- 21:24
- PR
- 21:25
- and see
- 21:26
- what I built and then committed to Grokbot
- 21:29
- And then
- 21:30
- I'm just using the same interface to chat with Grokbot
- 21:34
- and it's sending through
- 21:36
- all of my different workouts
- 21:38
- Under the hood, it's logging these in Git as commits to
- 21:41
- I believe, a JSON file
- 21:43
- But what I have is a piece of
- 21:45
- software with logic and a computer
- 21:48
- The agent is the computer
- 21:49
- and the interface is chat
- 21:51
- So while this might not look as beautiful
- 21:53
- or as full-featured as what I had before
- 21:56
- this is much more easily maintainable
- 21:59
- And if there's an error
- 22:00
- the bot is actually going to be able to diagnose that
- 22:03
- write code, commit to the GitHub repository
- 22:06
- and publish that
- 22:07
- And that's a much faster feedback loop than my training app
- 22:11
- right? Because in my training app
- 22:12
- I basically have to
- 22:14
- fire off cursor cloud agents
- 22:15
- I'd have to go home
- 22:16
- I'd have to, you know, maybe look at things
- 22:18
- approve, deny, merge
- 22:20
- Set up this like preview deploy feedback loop
- 22:23
- which is
- 22:23
- maybe a little bit tricky depending on the technologies that you
- 22:26
- that you use
- 22:27
- And so this is a fundamentally simpler approach to software
- 22:30
- whereby the feedback loop of
- 22:32
- iterating on that software is in the chat
- 22:34
- Another thing
- 22:35
- I kind of wanted to build AI chat
- 22:37
- into
- 22:38
- my
- 22:38
- training app. I wanted to be able to
- 22:39
- chat with it as if it was a coach
- 22:41
- With Grokbot
- 22:42
- I already have that out of the box
- 22:44
- So
- 22:45
- depending on what the software is that you're trying to build
- 22:48
- it might just be an agent
- 22:50
- And
- 22:50
- we're very soon going to ship multiplayer for
- 22:53
- tools like Grokbot. That's going to allow you
- 22:54
- to basically
- 22:56
- share the software with all your teammates
- 22:58
- So
- 22:59
- this is the personal software approach
- 23:02
- and something to keep in mind
- 23:04
- If you can break your app down into a client
- 23:06
- a server, And a database
- 23:09
- Consider attaching some sort of data source to Grok
- 23:12
- either by connecting it to Google Drive
- 23:14
- giving it access to a Git repo
- 23:16
- or just giving it some files
- 23:17
- It has a computer
- 23:19
- And then using the agent itself as the computer
- 23:22
- either by giving it skills
- 23:23
- and you can also just define skills for
- 23:25
- GrokBot. Something we didn't talk about
- 23:28
- You can attach files
- 23:28
- teach a task
- 23:29
- but then GrokBot also has access to skills
- 23:32
- if you define them
- 23:33
- MCP servers, tools, plugins
- 23:37
- and then using chat as the input and
- 23:39
- output
- 23:40
- That may be
- 23:41
- a replacement for a lot of the personal
- 23:43
- software or personal scripts you've been building
- 23:45
- Okay
- 23:46
- another way I've been using GrokBot is to write code
- 23:49
- Now
- 23:49
- we already talked about the
- 23:51
- sort of
- 23:51
- X bookmarks example where I'm building demos
- 23:54
- off of my Twitter bookmarks
- 23:56
- and I'm dispatching agents that way
- 23:58
- Now I want to talk about a bit
- 23:59
- more of an advanced concept
- 24:00
- which is doing actual software engineering
- 24:03
- from inside GrokBot
- 24:04
- And I think, Lauren
- 24:07
- per the cursor Really amazing engineer
- 24:10
- stated this best. I sat down with her at a conversation
- 24:13
- and she described Grokbot to me as the
- 24:14
- perfect outer loop. And so
- 24:16
- I actually call this agent outer loop
- 24:18
- Basically
- 24:19
- what we want is some outer loop where we're gathering context
- 24:22
- understanding problems
- 24:23
- pulling in
- 24:24
- context from these different systems before we kick
- 24:26
- it off to a cloud agent
- 24:28
- that
- 24:28
- acts as our inner loop and actually builds
- 24:30
- the software
- 24:31
- Why do we do this?
- 24:32
- Well
- 24:32
- we do it to protect the context of the inner loop
- 24:35
- the cloud agent
- 24:37
- and to give us a place to explore
- 24:39
- Why is this important? Well
- 24:41
- LLMs work on next token prediction
- 24:42
- right?
- 24:43
- That means that everything that's in context is
- 24:45
- going to influence the outcomes
- 24:47
- So if you're in an agent and you're like
- 24:49
- oh, read these files, what should I build
- 24:51
- should I build this
- 24:52
- help me think through these problems
- 24:53
- and then from that point you start delegating fixes
- 24:56
- that agent is actually going to have a
- 24:57
- bunch of dirty context and a bunch of
- 24:59
- information that might not be relevant
- 25:01
- to what you're trying to build
- 25:03
- So with this outer loop
- 25:04
- inner loop architecture
- 25:06
- we can protect the context of that inner
- 25:08
- loop
- 25:09
- and focus on collecting information with essentially
- 25:13
- an agent that's specialized for that
- 25:15
- So this agent I have here
- 25:17
- and If we look at the description
- 25:21
- this is an expert on the cursor codebase
- 25:22
- documentation and marketing repos
- 25:24
- so I gave it access to all of those things
- 25:26
- and its goal is to help me look
- 25:27
- through sources of context and come up with perfect prompts
- 25:31
- to then delegate to cloud agents
- 25:33
- with the skills that I use to build
- 25:35
- that cursor
- 25:36
- One of those is Lauren's
- 25:38
- PStack, plugin, which is amazing
- 25:40
- So an example prompt that I might send through here is
- 25:43
- do research on user credit grants
- 25:45
- This is a problem I've been thinking about
- 25:46
- Look across Notion and Slack and see how
- 25:48
- our company engages with granting user credits
- 25:50
- then come up with a plan to improve
- 25:51
- the system
- 25:52
- and kick off a prompt
- 25:53
- to a cloud agent
- 25:55
- So what this
- 25:55
- bot is going to do is going to
- 25:57
- go through all of those sources
- 25:58
- start to understand the problem
- 26:00
- Ideally, it would come back to me
- 26:01
- we'd chat through it
- 26:02
- and then the same thing we did
- 26:05
- down here with our tech demos
- 26:07
- it's going to fire off
- 26:09
- a cursor
- 26:10
- cloud agent to go out and actually build the thing
- 26:13
- Again
- 26:14
- the goal of this
- 26:15
- and why we might do it is to prioritize
- 26:17
- prioritize And protect the context of
- 26:23
- that inner loop coding agent
- 26:24
- and keep it more specialized
- 26:27
- So a bit of like meta discussion here
- 26:28
- a bit of future of building
- 26:30
- but I've been using this flow to basically kick off the
- 26:33
- the bots that I have working in cursor
- 26:36
- So a lot of my cursor
- 26:39
- bots end up started from Crockbot
- 26:42
- We're not replacing the coding
- 26:44
- coding harness. We're just augmenting it
- 26:46
- and staging for that coding harness
- 26:49
- Another really great use case
- 26:50
- I won't dive too deep on this one
- 26:52
- is just a cheap of staff or prioritization
- 26:54
- bot
- 26:55
- So I use this kind of for a catch-all
- 26:57
- like, hey, here are some really good links on DevRel
- 27:00
- can you put these in my Notion
- 27:01
- it'll just write to my database
- 27:03
- It also has access to my Gmail and my calendar
- 27:07
- so I can have it draft responses to emails
- 27:10
- which has helped me automate a few of
- 27:11
- those things
- 27:12
- And then I do a weekly review and
- 27:13
- a Monday weekly kickoff as well
- 27:16
- and it prompts me to do those because
- 27:17
- otherwise I just don't actually do it
- 27:19
- you know, so I have a bot keeping me accountable
- 27:22
- essentially. So that's another really great use there
- 27:24
- So we talked a little bit about coding
- 27:26
- we talked about personal software
- 27:27
- we talked about workflows
- 27:29
- and now I want to talk about information retrieval
- 27:31
- which is maybe one of the simpler ways to use GrokBot
- 27:34
- but
- 27:35
- honestly insanely helpful. And this is how
- 27:37
- I onboarded to Cursor
- 27:40
- If you want to see a write-up of this
- 27:42
- I have a
- 27:43
- Twitter article that I shared out
- 27:46
- not too long ago. We could go to X.com
- 27:50
- and
- 27:51
- the article is chat is all you need
- 27:52
- And I kind of talk about
- 27:54
- you know, lower on down here
- 27:58
- we have information retrieval
- 28:00
- I talk about how difficult it can be
- 28:03
- to access information across different resources
- 28:06
- You might have felt this if your company has a Notion
- 28:09
- a Slack, a different wiki, you know, you're responding to emails
- 28:12
- it can be really hard to find stuff
- 28:14
- And so
- 28:15
- my onboarding at Cursor was largely
- 28:18
- driven by Grokbot
- 28:20
- And I used
- 28:21
- actually some combination of this Cursor product expert bot
- 28:26
- number one, to answer questions about the product
- 28:28
- but then also to help me answer
- 28:31
- how the company works
- 28:32
- So you can even see when I was
- 28:33
- doing research for this For this
- 28:36
- video, I said, hey
- 28:37
- how do GrokBot group chats work?
- 28:38
- It has access to the code base
- 28:40
- It could just go and find that for
- 28:41
- me and tell me that
- 28:43
- all the agents that you add join the
- 28:45
- room. There's an orchestrator
- 28:46
- but it's not one of them
- 28:47
- And then
- 28:48
- the host runs a round robin after you send a message
- 28:50
- so it wakes each member in turn
- 28:52
- That's how I knew what to tell you all
- 28:53
- right?
- 28:54
- And I've been using this to learn about the product
- 28:57
- but also to search across Slack and Notion
- 28:59
- My onboarding really took
- 29:01
- like, a couple of days
- 29:04
- And it's still ongoing
- 29:06
- right? But I feel like I can answer
- 29:07
- all these questions
- 29:08
- because all of that information is accessible
- 29:12
- and saved off somewhere. All right
- 29:13
- so I've been talking a lot
- 29:14
- but the main things to remember here are that one
- 29:17
- Grok is an agent that's running in the
- 29:19
- cloud
- 29:20
- Grokbot is an agent with a computer
- 29:23
- and it can kind of do anything you
- 29:24
- can do
- 29:25
- So you really have to kind of expand
- 29:27
- what's possible here and think about writing code
- 29:29
- running code, running software, doing workflows
- 29:33
- controlling a computer the same way you can
- 29:35
- And if you delegate appropriately
- 29:38
- you're going to be able to accomplish quite
- 29:39
- a bit with this tool
- 29:41
- You can control what Grokbot is able to
- 29:42
- do or not do
- 29:43
- and you can access Grokbot through your phone
- 29:46
- or your desktop computer. It runs in the cloud
- 29:48
- so you can close your laptop
- 29:49
- You don't have to worry about it shutting
- 29:50
- off
- 29:51
- It's a persistent Linux VM
- 29:54
- under the hood
- 29:55
- on a server that's running in the cloud
- 29:57
- The main difference between Grokbot and many of
- 29:59
- these other personal agents you see out there
- 30:01
- are that there's zero setup
- 30:02
- and you don't have to do any server
- 30:04
- You don't have to do any configuration
- 30:06
- You can just jump in and start building
- 30:09
- and that's really impactful
- 30:10
- So I highly recommend giving it a shot
- 30:14
- Check it out. Let me know what you
- 30:14
- think. Let me know if this video helped
- 30:16
- you
- 30:16
- But until next time
- 30:18
- I'm Matt with SpaceX AI
- 30:20
- Peace
Watch these videos on your TV
XTTV brings video content from X to your living room. Download now and start watching on the big screen.