0:00

33:15

I Turned Claude Opus 4.7 Into a 24/7 Trader

Nate Herk | AI Automation·33:15·Technology·C1

Chương · 6

Chương 1·Building an AI Trading Agent with Claude Opus 4.7

Hiển thị

Auto-scroll

0:00

So, Claude Opus 4.6 is finally here, and it got me thinking because as I was scrolling down and I started looking at the benchmarks, I noticed right here we have a genetic financial analysis, and we have about a 4% jump from Opus 4.6.

0:12

And about a week ago, I dropped this video where me and Salmon traded stocks for 30 days with our open Claude agents, and my open Claude agent was set up with Opus 4.6, and I was able to beat the S&P by about 8% as you can see right here.

0:24

This green line is my chart, and this gray line is the S&P. So, if you guys want to check out this full video, I will link that right up here. So, after this challenge, I wanted to keep my agent running to see if it could keep beating the S&P, but now I want to upgrade it to 4.7 instead of 4.6. So, that's exactly what I want to cover with you guys today. How do we build a 24/7

0:42

AI trading agent with Opus 4.6 inside of Claude Code? We're not even going to touch open Claude or Hermes agent. We're going to just do this with Claude Code.

0:49

And this is all going to be thanks to another feature that we just got from the Anthropic team called routines. If you don't know what that is, I'm going to break it down in this video, but I also made a full one if you want to check that out, also right up here. So, basically this is the unlock. 4.7 plus routines means you can have a 24/7 AI agent, and 4.7 was built for full

1:06

throttle agentic work, judgment over ambiguity, and self-verifying outputs. So, it's perfect for these types of routines, and that's how I'm going to set up my 24/7 trading agent with Claude Code routines. So, if you follow along with this video, you'll be able to pretty much duplicate my exact setup and have yourself a working agent and start trading with it. So, let's go ahead and

1:25

get started. What are we building? We're going to create a Claude Code project that runs on a schedule. Something like a pre-market cron, a market open, a midday, and a close. It's going to research the market. It's going to place trades via the Alpaca API. It's going to journal and make sure that it's writing context into the file so it can learn, and then it's going to send me every

1:43

day, every day of the week, an end-of-day summary to my ClickUp. The goal, obviously, is just to beat the S&P. We're not looking to do any crazy crypto or crazy day trading stuff. I'm just kind of doing this as a fun long-term investment. Just beat the S&P challenge. Like I said, in my first 30 days, I gave Opus $10,000, and we were able to beat the S&P by about 8%, which

2:02

is really good, especially in just such a short amount of time. So, what is the tech stack? We're going to be using Claude Code routines. That's kind of our scheduler. For the AI model, we're going to use Opus 4.7. For the skills, we're going to be using just kind of like building our own as we go. It'll be stuff like how the agent does research or how it makes decisions or how it

2:19

places trades or, you know, makes logs, things like that. We're going to use the Alpaca API as our brokerage. We're going to use the Perplexity API for research.

2:27

You could also just use the native web fetch and web search, but I'm going to use Perplexity here, and then I'm going to get my notifications in ClickUp this time rather than Telegram, just cuz I live in ClickUp a lot more. So, obviously, some of these things can be interchangeable for you, but if you want to just copy my exact setup in this video, then feel free. Okay, so how to

2:44

get started with these different tools that we need for this video? The first one is going to be Alpaca. So, go to alpaca.markets.

2:50

You can see up at the top we even see want to trade using AI. We launched an Alpaca's MCP server to help you trade with natural language. So, you could use that. I'm just using the endpoints. I honestly like to use endpoints a lot more. But anyways, you're going to sign up right here, and you'll click on trading API. Now, once you get in there, you will have basically a paper trading

3:06

account, and right here you can see I've got a paper trading account with 100K. You could also open up a new account right here, and that's how you could put real money in. You might, like I said though, have to just verify your account, which could take a couple days.

3:17

And then they're both the same as far as the credentials. You basically scroll down, and on the right-hand side you have your API keys right here. You get an endpoint, and you get a key, but you also need like a secret key. So, you have like a secret key, and then you need like an ID. So, you'd hit regenerate, generate new keys, and this is where you get your key and your secret. And

3:33

once you copy and paste this, just make sure to save it because then it's going to go away, and you'll only be able to see the key, but you need both of them.

3:40

Next up, I have ClickUp. This is where I want to get my notifications. You don't have to use ClickUp. You can use Slack.

3:45

You could use, you know, whatever, Telegram, WhatsApp. I would go up here and go to my settings, and then I scroll down and find ClickUp API, and then I would just copy this token right here.

3:54

And then similarly with Perplexity, I would go down here, and I would click on all settings, I think. And then in here I can see API platform, and this is where I would be able to go find my actual API key. If I went down to right here, API keys, I could make a new one or copy an old one. But basically, the way it works is every platform has typically API keys. You just have to

4:13

find them in the settings somewhere. If you can't find them, just do a quick Perplexity or Google search or even ask Claude Code, and it should be able to help you get to your API keys. And now that you know how to get these keys, just hold on to them. We're going to actually plug them in later in the video. And then the final piece of the puzzle is going to be the Claude desktop

4:27

app. You can see here you can have Claude chat, you can have Claude Code work, or you can be in Claude Code. And this is where we're going to set up these actual routines, where now if I go to my calendar, you can see every single day I've got Bull, which is my trading agent, waking up at 6:00, 8:30, um noon, 3:00 p.m. And then on certain days, like on Fridays, he's doing a weekly review.

4:47

So, this is the way I currently have mine set up, and we're able to manage everything right from here inside of the Claude desktop app because of this whole routine feature. So, all you need to do to use the Claude desktop app is just type in Claude desktop app download, and you should be able to install it right here. You go to the downloads page, and then you just have to choose the Claude

5:03

desktop for your operating system. And then the only other thing you need to do is make sure that you are on an actual subscription. So, it has to be a paid, either 20 bucks a month plan or the max plans. Now, in this video, you might also see me use Claude Code inside of something that looks like this, which is the VS Code IDE, and then we use Claude Code extension inside of it. It's not

5:22

mandatory, but that is where I typically prefer to use Claude Code because I can see all of my files over here. So, if you want to do that, just type in VS Code in your browser, open this up. It's completely free to download, and then go ahead and download it and connect your Claude Code account. So, that is pretty much the tech stack that we're going to be using today to get all of this set

5:38

up, and let's get back to the video. Okay, so before we get into the actual step-by-step kind of build, we are going to need to talk about the mental model.

5:46

So, understand this before you write anything. The trading strategy is a piece of the work, but the memory architecture that you're going to use is going to be huge.

5:55

So, every time a routine fires at 7:00 a.m., Claude Code basically wakes up essentially stateless. It doesn't really know anything. So, how do you make a stateless agent act disciplined and remember rules and learn over time? You do that with files and with context. So, every routine basically wakes up, reads files, does the job, and then writes back any important lessons or anything in like

6:17

that the next agent when it wakes up needs to know, it just writes that down. And that's kind of at a very high level the type of framework that we're going to be using here. So, the final thing I wanted to talk about before we get into this is the context budget. So, treat tokens like money.

6:30

Obviously, every file that you ask the agent to read, it's going to cost you. Each routine gets about 200,000 tokens to work with. I mean, we do have the a million token context window models now, but I don't really like to think of it like I have a million tokens to play with because context rot. So, system instructions will take some up, strategy files, your trade log, your API, your

6:48

research. Every run is going to take a different amount of tokens, but just keep that in mind. And the reason here that we're using, if I go back to the stack, Claude Code instead of just using something like, you know, the agent's SDK or I mean, I guess that could work, but rather than just using like a standard automation, is because if we do this with routines, we basically get the

7:06

full autonomy of us working with Claude Code manually because it goes through that whole agentic earnest loop of figuring things out. And especially with Opus 4.7, it's apparently going to be really, really good.

7:18

Whereas, if you just had like an automation that would do research and then, you know, make a trade, you're not getting that super variability that you probably want with something that could be very volatile like trading. All right, so now that we got all of that stuff out of the way, let's get into the strategy. So, step one, think of this like you're teaching a kid

7:36

to ride a bike. You have to explain how to balance before you go. You have to explain to the kid that he's going to steer with his hands, and he's going to pedal with his feet. And even once you put him on the bike, you're not just going to let him go. You're going to hold the handlebars. You're going to put your hand on his back, and you're going to walk with him until he gets a little

7:52

more comfortable. And then slowly, you might take off, you know, you might stop steering, but you might just run behind him. And then slowly, you'll put on some training wheels. And like you just keep bumping it up in phases. You can't expect to just throw a kid on a bike, and it's instantly going to be magic.

8:05

So, maybe you want to start with paper trading. This is obviously not financial advice, and I'm not, you know, recommending that you go put all of your life savings in here and expect to make money and then get mad at me. This is an experiment for me, and I'm trying to show you guys what's possible. So, just keep that in mind. Start with paper trading first if you're not comfortable.

8:21

But anyways, the idea is you need a strategy. So, for me, what I'm going to do is I'm basically going to extract all of the strategy that we've been using from my open Claude agent, and I'm just going to say, "Hey, let's take this over to Claude Code." So, for me, it's a little easier. But for you, if you actually are like a manual trader, and every day you think about things and you

8:37

do research and you place trades or you sell things, then just write down your strategy. How often are you checking the news? What are the signals that make you realize I'm going to buy this or I'm going to sell this? As much of your gut intuition and your routine about trading that you can get into this Claude Code project, the better. And then I over time, you are going to help it learn

8:55

from its mistakes. You're going to iterate, and it's going to get better and better every time you use it. But if you're starting completely fresh, which in my case of the open Claude trading challenge, I literally just said, "Hey, I'm doing a 30-day challenge. I'm trying to compete against my friend Salmon. Go do research. Spin up a team of sub-agents. You are my wealth advisor.

9:13

Try to beat the S&P. Figure out the best plan possible." And that's like genuinely all that I said to it, which I think is pretty cool. And it went off, and it did a bunch of research. It figured out a strategy for me, and then for the next 30 days, it just did that, and I barely had to touch it at all.

9:26

So, if you want to start off by playing around with some research like that, then certainly do that. But if you guys remember at the beginning of this video, I kind of alluded to this benchmark here. So, in the official release, and by the way, I just dropped a full video on this new release. I'll tag that up here. So, in the official release of Claude Opus 4.7, they always drop the

9:43

Benchmarks, and they compare it to other models that are kind of leading in the industry as well. And so, this is the one I wanted to talk about. In this case, we have agentic financial analysis 64.4%.

9:54

We also have stuff like gradual level reasoning, and you know, just the ability to think really well. Agentic search actually is worse than Opus 4.6, which is interesting. But, yeah, Agentic financial analysis. So, what does that actually mean? This basically means that it is really good at analyzing a company, but it's not basically a signal or a benchmark for how good this model

10:13

is at trading and specifically day trading. Like this doesn't mean that it's going to be able to look at like a technical analysis and look at the candlesticks and the MACD and figure out the exact best moment to buy and then the exact moment to sell. I'm not going to try to day trade with these agents. I just don't know enough about it to do it. But, that's not what this metric is

10:29

saying. This benchmark rewards models that can digest filings and write coherent fundamentals driven thesis or thesis, I guess. And that maps to long-term or swing or fundamentals driven strategies, not day trading. So, that's why once again, we're just going to be trying to beat the S&P here. Okay, so that is the overall strategy. Let's talk about the scaffold. So, what you

10:49

need to do to spin up the Claude code project in about 60 seconds is you're going to open up a new folder and we're going to open up Claude code. So, let me hop over to my project right here. This is a new folder. It's called trading I think trading routine. This is basically just a settings file. There's hardly anything in there. The claude.md is blank and then we have a read me just

11:07

because I turned this into a GitHub repo so I can pull it into a routine later. If you haven't done that yet, it's fine.

11:12

We'll talk about that in a sec. The reason why we're starting here in VS Code is because I want to get started in a place where I can see all of my files on this left-hand side because if you're in the desktop app, which we will move over to here later, you don't get to see all your files open on the left-hand side. On the left-hand side, you see your routines and your other sessions,

11:28

which is great, but when we're first getting set up, I like to be able to see everything in one spot. So, we're going to start here in the VS Code either terminal or the Claude code extension.

11:38

So, let me show you guys what I did. This is my Telegram chat with Bull, my trading bot, and I said, "Hey, I want to completely migrate you to a different AI agent setup. I need you to break down the strategy, the cron jobs, what are you looking for when you're doing your research, what type of sub agents do you have, what are the different signals, give me the full breakdown of all of

11:55

your learnings." And so, it gave me all of this stuff. And so, what I'm going to do is I'm just going to basically migrate over this strategy into my Claude code project. You can see we have all of these cell signals, we have all of these key learnings and these kind of main frameworks and we have our current portfolio. Real quick, guys, when it comes time for you to get set up, I have

12:14

this free resource for you. It's a 13-page PDF and I am assuming if you dropped this thing into Claude code and then started a brainstorm with it, it would help you out a lot as far as the infrastructure and the folder structure and the way you want to get things set up and the way you want to create your routines and things like that. So, it should be very helpful. This you can

12:30

download for completely free in my free school community. The link for that is down in the description. Once you get in there, just head over to the classroom, click on YouTube resources and you will be able to find it right in there. So, let's get back to the video. So, I pulled it up real quick. You can see here's today. We had a bit of a jump and then we came back down today. Here

12:45

is the past month. So, this is where we were with the open claw agent. But, you guys are going to want to go to Alpaca and make an account, especially if you want to put real money in. You might have to take a few days for your account to get approved. So, head over there and get an account. Then I also told it to give me a technical cheat sheet as far as like the different API

13:01

end points that uses and how it makes those requests, as you can see, just because it already has all this knowledge. So, why waste time trying to train a new agent if we could just migrate over all of these memories and knowledge to this new agent. And at the end of that, it even gave me an export of all of these important files like the agent instructions, the trading

13:19

strategy, the trade log, the research log, the weekly review. And I'm just basically going to migrate all this over. So, let me show you what that's going to look like. So, you can see that it gave me this zip file of all of those seven files and I'm just basically going to extract those into this project right here. So, I'm choosing the trading routine project, select folder, and then

13:36

extract and we should see all those files just popped up right over here on this left-hand side. Now, what I'm going to do is just talk to my Claude. I'm going to switch over here to plan mode and I'm just going to start talking. All right, so I am creating you as a 24/7 AI agent trade bot for me. I had a project set up with Open Claw and we were trading, you know, $10,000 using

13:57

Alpaca for the past 30 days and we were doing pretty well. But, now I want to move it over here so we can set you up as a Claude code routine. So, I just dropped in a ton of context for you.

14:06

There's a ton of files in this project. If you want to just take a second to get sort of like acclimated to what I dropped in there.

14:13

There's going to be a lot of information, but basically that's like all of the knowledge and all of the strategies and all of the signals that our open claw agent had and how it was able to be successful and I just want to migrate that over here to you so that we're not losing any of our knowledge. I'm also going to paste in some of the key things that it told me

14:29

that you should make sure to know. So, I'm going to drop in that stuff right below. Now, some of this stuff you can take with a grain of salt because it might have been set up and optimized for an open claw agent harness, but you're Claude code. You understand the best way to set up your projects and to do things. So, what's really important is that you're ingesting this information

14:47

And you're understanding it and then you're helping yourself organize this project in a way that makes the most sense to you. Okay, so I know that is a ton of information that I just chucked into there and you guys might be sitting there like, "Okay, well, I don't have all these files. What do I do?" This is where you just start to brainstorm. Literally treat this thing

15:03

as if it is your best friend who is the best trader in the world and you are just trying to give it your strategy now. So, if we think back to earlier in step one where I talked about the strategy, when you open up a piece of paper or sorry, you take out a piece of paper, you open up a Google Doc, and you just start writing down.

15:17

How do you actually trade? What's your approach? Once you write down all this stuff and you basically just brain dump about how you want to do it, then you plug it into the Claude code and that's where you see how I set up this project here. I'm just helping it get situated with what we want to do right now. And honestly, the more context and details that you give it right now, the more it

15:35

asks you questions, the better. As you can see, I'm using the brainstorming skill from superpowers and it's going to start off by asking me some questions.

15:42

Okay, so I just answered some questions and I did want to show you guys this. So, it alerted me about something with security. Inside of this template that Open Claw gave me, it had my live Alpaca key in there. So, I'm going to go ahead and rotate those before push this new one into production, but just something to be thinking about when you're migrating stuff or when you're feeding

16:02

in API keys directly to a chat. Just be careful of that kind of stuff. And I'm going to show you how we get our API keys set up in routines because you don't want to put them straight into a GitHub repo. You're going to have to set them up in the environment variables, so I will show you guys that as well. Okay, so this point it's come back with a plan. So, Nate ran a live money Alpaca

16:17

trading bot called Bull on the Open Claw harness for 30 days and now we have all these context files with strategy, with trade history, with research notes, blah blah. The goal of this migration is to make Claude code a first-class home for this bot. So, we will set up the cron scheduling, the notifications.

16:32

We're getting rid of that agent mail step. So, me and Simon were having our agents trash talk each other. We're going to rotate our credentials, of course. And then it's proposing a layout for the way that the files should be set up. So, you can see it's going to put the files in here into the memory and then we're going to be able to prompt our routines to always check the memory

16:48

first and update the memory. So, this is where it will really help for you guys to brainstorm with how the project should be set up. Oh, one other thing that I forgot to give you here is that I wanted to use Perplexity for research.

16:58

So, I'm going to come down here to the bottom of the plan and I'm going to say, "Yeah, this all looks pretty much good to go, but one thing I forgot to tell you is that we want to use Perplexity for research rather than just the native Claude code web search and web fetch functions. So, add in that Perplexity will be used for web search and once again, once we set up the routine,

17:15

that's when I'll give you the Perplexity API key." Okay, so after this change has been made, we should be ready to start migrating over to setting up the routines. All right, so the new plan has come back. So, here's the end-to-end check, secrets, Alpaca, memory files/commands, dry run, ClickUp, Perplexity, nice. We're going to go ahead and accept these changes. And they

17:35

also noticed that we have auto mode on. So, I think with this new update with Opus 4.7, when you auto accept a plan, it turns on auto mode, which means it lets Claude handle permission prompts automatically. Now, you will have to be careful though because doing this is a little bit more expensive. So, if you don't want it to be auto mode, then you can just switch it right off of auto

17:52

mode, of course. But, um yeah, just something I wanted to point out. All right, so while all of that's still getting set up, let's talk about step three, which is guardrails.

18:00

So, before you start working in the trading logic, think about this kind of stuff. I obviously talked about how you could start with paper mode and then you toggle on real trading and put in money when you feel comfortable. You could do things like, "Hey, max 5% of my portfolio per position." You can have a daily loss cap. You could tell the agent what not to do. Like only buy three new

18:18

positions per week or no options ever or don't do this or always do this because the agent is going to be autonomous. It's going to be eager and if you don't give it guardrails, it might just kind of start to, you know, go off the rails.

18:28

But, like I said, even though I'm migrating a system over, I know it's not going to be perfect. So, I'm going to watch every single run and I'm going to read through the conversation history of every run and I'm going to see if there's things that I need to tweak in the prompt or things that I need to tweak in the project settings or the context files. And like I said, just

18:44

kind of keep building the plane as we're flying it here. And then that's where we also start to develop skills because right now we have skills maybe for like a research skill or for a trade skill, but we're going to make sure that we're getting them actually explicitly invoked and we're going to be making sure that that increases the consistency and we're going to make sure that the agent is

19:01

able to give us insights every day of, "Hey, here is what worked well and here's what I noticed and hey, here's where I think that maybe we should update these skills X, Y, and Z, you know, whatever." So, a combination of thinking about the guardrails and thinking about skills is how you can continuously make this thing better over time. And now that we've talked about

19:18

all that, we will get to the routines and the deployment. Let's check in on our guy right here and because I turned it off auto mode, it was sitting and waiting for me to do a permission. All right, so everything has been migrated.

19:30

We're looking good, but we're getting a little bit to the point where I like to reset the context. So, before we move on, I'm going to say, "Hey, give me a summary of what we've done." And then I'm going to clear the context and we'll get started again. All right, so here's the session summary that it gave me.

19:43

Copy that. Do a /clear. Paste that in, hit enter, and now we're basically right back at the beginning with a fresh window of context. So, this is where it gets pretty cool. You can see on the left-hand side it has reorganized things. We have our scripts, we have our memory files, we even have some commands in here, and it built all of that just because I told it that I

20:01

was trying to migrate stuff over. It also updated this clawed.md file because now we have bull back in all of his glory, and we have our clawed.md file that is about 156 lines. So, now it's time to start planning out step five, which is the routines. So, we basically need to figure out what are the different daily triggers that we actually want. And we also don't maybe

20:20

want them every day because the market's only open 5 days a week. So, this is kind of like a baseline. This is kind of how it worked with my bull trading agent, but I'm actually just going to come into VS Code, and we're just going to brainstorm once again. So, I'm going to go on plan mode, and I'm just going to start asking. All right, so now that you know everything about this project,

20:37

and you understand what we did with open claw, I want you to help me figure out what is the best way that we should be setting up these clawed code routines to basically simulate that exact same type of environment. So, we have those cron jobs, but each one is kind of doing a bit of a different thing. You know, the pre-market one does a bit of a different job than the mid-market one, and the

20:56

day close one does a bit of a different job. So, you have a better idea of all of our memory files and our skills, and you understand the structure of every time the agent wakes up, and the routine is triggered, they need to read the files first to, you know, kind of get it acclimated with the environment, do what they need to do, research, potentially make a trade, potentially sell

21:15

something, you know, whatever it is based on their best judgment, and then update all of the memory files so that next time the next agent wakes up, they understand what's going on. So, what I want you to do is help me figure out what are the right structure for these um cron jobs, and most likely we only want them during the week when the market's actually open.

21:35

And just like that, I'm able to brain dump, and Claude's going to do the heavy lifting for me and propose some ideas. Okay, let's see the plan that it came up with. So, stand up five recurring clawed code scheduled triggers, and we've got these different slash commands that it's going to be able to utilize. So, all five triggers will run on the weekdays only except for the weekly review, which

21:53

is Friday only. So, we have a pre-market, which is for research catalysts and draft trade ideas. It won't message me unless it's urgent. We have a market open, which is going to execute the planned trades and set 10% trailing stops that will go to the trade log. Only if a trade is placed will I get a notification. We have midday, which will cut -7% losers and tighten

22:13

stops on winners, also go into trade log. So, here you guys can see basically it was able to analyze our strategy and analyze what we have access to, and then build me a plan. And now what I want to do is basically just have it create those prompts for me so that I can go ahead and go set up those routines in clawed code. All right, so those crons look great. So, I don't need

22:32

you to actually create them. What I need you to create for me is the prompts. As if I was going to run one of those commands into a fresh clawed code session, and I wanted it to do all the right stuff, what is the exact prompt that I would enter into that session?

22:45

And the one thing that you need to keep in mind here is reminding it to obviously um read the files before it does anything, and update the files at the end, and then also when it needs to use an API like Perplexity or ClickUp or Alpaca, that it should be grabbing those keys from the environment variables of that environment, not inside a .env file. All right, and now with that

23:05

feedback it should be able to give us those prompts, and then we should be able to just set up the routines and go ahead and test them out. Okay, so while this is writing up, there is one more thing that I wanted to show you guys that is very important. So, in the clawed code desktop app, when you go to routines and you go to create a new one, you can see that you can create them

23:20

locally or remote. So, if you create them locally like some of these I have, this just basically means that it lives in your local directory. It can still talk to the internet and do all those things, but it lives in your machine essentially. And that means if you closed out of this desktop app, that local routine would not fire off. And that's why these remote routines are so

23:40

cool because your computer can be off, but they're still running because they run in the cloud. But what that means is when you do a remote run, you have to have a GitHub repository to have this running, which is why I have this trading repo set up.

23:54

And so when they run via a cloud, you know, routine, it basically clones that repo in the cloud, it spins up a copy, it works out of it, it does what it needs to do, and then it basically like destroys that cloud environment. So, if you want them to run remotely, and you want those files to persist every session, you just have to talk to clawed code and say, "Hey, make sure, you know,

24:12

because we're doing a routine, make sure that all of these files that it's actually updating, it's able to push back and commit back to that main actual branch." Back to that main, and we just got a notification from my trading bot because it was market close. But anyways, I was saying, make sure that it's able to push those changes back to the main repo, otherwise the next

24:33

agent's not going to pick it up, and then what's even the point? So, you can start local if you want, and then you can um sort of graduate to the cloud if you want, but I definitely just wanted to make sure you guys understood that difference there. Okay, so that's done.

24:44

You can see it created this new folder right here called routines, and this is basically the prompt that we'd be feeding into each of the different routines that we make in order to make sure that bull actually understands what to do. And the last thing we need to do is actually push all of these new changes to this GitHub repo because right now here is my trading routine

25:02

private repo. All that's in there is a readme, and I don't think there's anything in here. So, we need to get all of these files that we just created, you know, these commands, these scripts, these memory files. We just have to get all of this into that repo so that our routine can actually access it. So, all I have to do is say, "Hey, push all of these changes, commit

25:17

everything to the repo." By the way, if you've never done that before, it's super simple. You go to github.com, create an account, it's free to create an account, and GitHub is really just like um it's like working on um a Microsoft Word on your computer, and that's like local. And if you wanted to share it with someone, you would have to put it into like the shared drive or, you know,

25:36

A Google Drive or something. It just means that you could grab a code base from a different machine because you've uploaded it somewhere on the cloud. So, that's kind of why we're doing this whole thing with the routines in the cloud. So, when you get to GitHub, you come here, you create a new repository, or the easiest way to do it is you just come to your clawed code, and you say,

25:53

"Hey, help me turn this into a GitHub repo." And it will say, "Okay, I need to authenticate you. Can I use the CLI?" And you say, "Yes, go ahead and do that." And then it will basically open up a tab, and it will say, "Hey, can you sign in with your GitHub?" You sign in, and now your clawed code just has your GitHub credentials, and whenever you need to push changes or clone a repo or

26:11

whatever you need to do with GitHub, it can just take care of it because GitHub's been around for a long time, so it knows it way better than you probably do. It knows it way better than I do.

26:19

So, when in doubt, just ask clawed code. All right, so if I give this puppy a refresh, we should see we now have all of these files popped in here. We've got bull trading routine. Awesome. So, let's head over to the clawed desktop app now and start getting set up. So, what you're going to do is you're going to click on new session, and you want to open up the local project right here of

26:38

whichever thing you just set up there. So, right here you can see that I'm in the folder called trading routine, which we were just working in VS Code. And if you guys remember, in VS Code we had these five different routine prompts, and inside of each of these they had the cron, they had the actual like steps, and then they had like any of the skills or commands that they need to run. And

26:57

this is in a folder once again called routines. So, before we start talking to our new project, we need to set up an actual environment for this. So, if I go to new routine and I go to remote, and I go over here to my um cloud environments, you can see that you can add one. And so, just add one, and I'm going to pull up my trading one so you guys can see. But here, this is where we

27:16

set up things like our API keys and our network access. So, I called mine trading, I gave it full network access, and this is where I'm going to put in my API keys for Alpaca and for Perplexity and for ClickUp.

27:27

And then you basically just go ahead and save the changes of that environment, and this is the one that you want your actual routines to be scheduled inside of. I am going to basically just talk to it right here to help us get these remote routines set up. Look inside the routines folder inside of this project, and help me set up all of these different scheduled routine runs. They

27:49

should all have a cron job in there. They should all have a prompt in there already, and we will be working out of this GitHub repo and scheduling off these routines. So, let me know if you have any questions. Do not exit the planning phase until you're 95% confident that you know exactly what to do and exactly what I'm looking for. So, now I'll shoot that off, and it should

28:06

be able to come back with a plan to actually go create all of those cron jobs for us. Okay, now you can see it's coming back and asking us some questions. So, the first one is which cloud environment should they run in?

28:15

We're going to do trading. That's the one you guys saw me set up. We want to use clawed opus 4.7, and we're going to use this GitHub repo. Yes, trading routine. All right, so it's set up all of those routines for us. You see that all I had to do was talk to it. It set up these crons cuz I don't know what exactly this means, but basically that means 6:00 a.m. Monday through Friday

28:33

for all of these. Monday through Friday 8:30 a.m., noon, 3:00 p.m., and then this one is just 4:00 p.m. on Friday. So, if I go over here to my routines, and I scroll down, you can see that we have these new five routines for bull, and these are all remote. If I click in this one, for example, says every Friday at 4:00 p.m. Central, we're in the right repo. If I click into here, we should

28:51

see we see this entire prompt, which is correct, and we're in the right environment, which is my trading cloud environment. And then one other important thing that you have to do is inside of your actual um routine, what you want to do is go to permissions, and you want to turn on allow unrestricted branch pushes. This removes the branch push restrictions, and clawed will be

29:10

able to actually push to any branch, not just the clawed slash branches. So, turn that on for all five of these routines here. So, I'm going to turn that on for all five of these routines here. So, let's just pretend that it is Friday.

29:21

I'm going to do a run now to see if this is able to actually work. And when you run these, you can actually click right here, and you can watch them actually work in real time. And best practice with routines is that every single time you build a new one, you should always do a run now at least a few times to make sure that they are working as expected. And exactly why we test out.

29:37

You can see what happened is when it tried to hit Alpaca, it wasn't able to find the API key because for some reason, once again, it thought it was going to be in a .env, but as you guys know, we don't have a .env because we don't put that in the GitHub. So, I'm coming back into this session that's set up these five and I'm just going to say, "Hey, remember all of the API keys are

29:53

going to be stored in the environment variables. So, all of the prompts must explicitly state that." Okay, so the issue was they weren't spelled exactly word for word, letter for letter. So, I fixed that and you can even see, for example, this one. They're all now coming through, but for example, in this one, I have the word API right here between Alpaca and secret. So, they have

30:11

to be matching one for one, otherwise it will think that they're not there. But anyways, now that has been figured out, it's actually able to look through all of these things, it's able to look through the memory, and it's going to execute our weekly review here. And there we go, just like that, the weekly review has been ran. I also told it that I fixed the Alpaca secret key, so now

30:27

all of that's been updated. And if you go over to my GitHub, you can see that it was able to push these actual commits to this repo. This one was just a test one that it was running on the background as well, but it was able to actually give us the weekly review and you can see it was committed by Claude and it made changes to this actual MD file. So, now the next time it actually

30:46

will be able to pick it up and we have our stuff all synced up. And alongside that weekly review, you can see that in my internal automations channel, sandwiched between two YouTube notifications, we have the weekly review that the run just sent us, we have our portfolio, we have versus the S&P, and we have any trades that we were making, and then, you know, some best and worst,

31:05

And also it graded itself a C for the week. So, we'll see how this is able to improve. So, I just spun up this test so you guys could see that we are able to talk to Alpaca right now. You can see that we're able to get the current balance. And if I go into here, we should see 9859 and back in Claude desktop where it pulled it, 9859. And just like that, we have set up a 24/7 AI

31:26

agent trade system with Opus 4.7. We've got our 6:00 a.m., our 8:30 a.m., we've got our noon and our 3:00 p.m., and then on Fridays we have our actual weekly review as well. Now, of course, like I talked about earlier, it is on you to now keep testing, keep iterating, making sure that everything's working, and upgrading your strategy as you go. And maybe you want to add some more crons

31:44

in, maybe you want to take some away. Maybe you want to add one on just like a Sunday night to do some more research to prep for the week. However you would actually want to set up your trading strategy if you were doing it manually.

31:54

Now, I know you guys are thinking like I have four automations per day basically plugged in now with Claude routines and that's going to eat away at my session limits, right? I mean, yes and no. I'm definitely going to be watching it, obviously, but keep in mind when I did this with my open Claude agent for 30 days, I was also plugged into my Claude subscription and I was still doing my

32:12

own workflows and I was never really hitting my limits. It's all about managing your context. And I'm going to have a full video dropping about that, another one coming out soon. But I can tell you for certain that if I were to set up all of these crons and have the same sort of autonomy and agentic workflowness, and I was just doing this via API, then that would 100% be way

32:31

more expensive than what I've got set up here. We have now deployed and now we just go ahead and track. And what is the hidden lesson here? The hidden lesson is that files aren't just memory, but they're essentially the agent's full personality and discipline. Like with open Claude, you set up your soldat.md file and you've got your agents.md file and you've got all these different

32:46

files. So, it's really about how do you make sure that you orchestrate this in the right way so that it can access the right ones when it needs to. So, anyways, I hope by the end of this video you guys are able to follow along and set this up. I'm going to be giving away this document and it's going to have way more information in there basically about everything that we've gone over

33:01

today and I'll drop that in my free school community. The link for that is down the description, once again, completely free. But that is going to do it for today. So, if you enjoyed the video or you learned something new, please give it a like, it helps me out a ton. And as always, I appreciate you guys making it to the end and I'll see you in the next one.

33:15

Thanks, everyone.

Transcript

Hiển thị

Building an AI Trading Agent with Claude Opus 4.7

The speaker introduces the concept of building a 24/7 AI trading agent using Claude Code routines and Opus 4.7. He explains that he beat the S&P by about 8% in a 30-day challenge using Opus 4.6, and now aims to improve with Opus 4.7. The agent will use Alpaca API, Perplexity API, and ClickUp for notifications. The mental model and memory architecture are emphasized as crucial for the trading strategy.