Claude + GitHub = The Most OVERPOWERED AI Trader I've Ever Seen — Transcript

Full transcript

1. 0:00In today's video, I'm going to give
2. 0:01$100,000 to a team of 12 AI agents to
3. 0:05trade on my behalf. And this is using a
4. 0:07free GitHub repo called Trading Agents
5. 0:10with 107,000 stars. Over the past few
6. 0:12weeks, I've been experimenting with
7. 0:14GitHub repos to see what the best
8. 0:16automated trading tools on the internet
9. 0:18are. And these are repositories created
10. 0:19by real developers. As you can see,
11. 0:21Trading Agents has 21 contributors and
12. 0:24is getting constant updates by real
13. 0:26developers. So, the good thing about
14. 0:28using GitHub to source your tooling is
15. 0:30that you don't need to build it from
16. 0:32scratch yourself. You've literally got
17. 0:33pro developers that are building this
18. 0:35stuff for you. By the end of today's
19. 0:37video, you're going to know how to have
20. 0:3812 trading agents trading on your
21. 0:40behalf. It's literally like having an
22. 0:41entire team of quant traders. Who would
23. 0:43have imagined a few years ago when I
24. 0:45first started AI trading, getting pine
25. 0:47scripts off GPT, and putting them into
26. 0:49TradingView, that you could literally
27. 0:51now spin up an entire army of AI agents.
28. 0:53Well, we are now in that era, and AI
29. 0:55trading is only going to accelerate. So,
30. 0:57the sooner you can learn it, the better
31. 0:59positioned you are going to be. So, if
32. 1:00you want to learn more about AI trading,
33. 1:02subscribe to the channel, and let's get
34. 1:04into the video. So, firstly, before we
35. 1:05start trading, let me walk you through
36. 1:07Trading Agents and explain what it does.
37. 1:09So, essentially, it's a multi-agent
38. 1:11trading framework that mirrors the
39. 1:13dynamics of real-world trading firms.
40. 1:15So, it's basically like having your own
41. 1:17AI trading firm. But applying
42. 1:19specialized LLM-powered agents from
43. 1:21fundamental analysts, sentiment experts,
44. 1:23technical analysts to traders, a risk
45. 1:25management team, the platform
46. 1:26collaboratively evaluates market
47. 1:28conditions and informs trading
48. 1:30decisions. These agents engage in
49. 1:32dynamic discussions to pinpoint the
50. 1:34optimal strategy. So, just like in a
51. 1:35real trading firm, you would have an
52. 1:37analyst presenting data from the real
53. 1:39world to the risk management division,
54. 1:40and the risk management division would
55. 1:42evaluate sizing, and then it would get
56. 1:43ticked off by the head trader. You're
57. 1:45essentially mimicking that process with
58. 1:47a series of AI agents. So, you can see
59. 1:49here. So, this is exactly how it works.
60. 1:51The repo plugs in the market data from
61. 1:53Yahoo Finance and other data sources
62. 1:55into the researcher team, also social
63. 1:57media trends, so X and Reddit. Then
64. 1:59you've got all of the new sources like
65. 2:00Bloomberg, Reuters, and then it taps
66. 2:02into the fundamentals so a company
67. 2:03profile, financial history, insider
68. 2:05transactions. All of this information is
69. 2:07then routed to the researcher team which
70. 2:10decides whether an asset is bullish or
71. 2:12bearish. It provides buy evidence, it
72. 2:13provides sell evidence. Then the trader
73. 2:15makes a transaction proposal and the
74. 2:17risk management team decides whether to
75. 2:19be aggressive, neutral, or conservative.
76. 2:20That then goes to the head manager and
77. 2:22then they are responsible for making a
78. 2:24decision on execution. And obviously in
79. 2:26your case if you wanted to manually
80. 2:27approve you could. If you want to leave
81. 2:29things fully to the AI then you can also
82. 2:31do that as well. So it's essentially an
83. 2:32entire team that is trading on your
84. 2:34behalf. So here are the 12 agents that
85. 2:37you get access to. You've got the
86. 2:38analyst team who's the fundamental
87. 2:39analyst, you've got the sentiment
88. 2:40analyst, a news analyst, and a technical
89. 2:42analyst each with their own personality,
90. 2:45profile, and skills in order to conduct
91. 2:47each individual role. Then you've got
92. 2:48the researcher team who critically
93. 2:50assess the insights provided by the
94. 2:51analyst team. Then you've got the trader
95. 2:53agent which composes reports from the
96. 2:55analysts and researchers to make
97. 2:57informed trading decisions and they'll
98. 2:58give a reasoning and a recommendation
99. 3:00for acting in a certain way. And then
100. 3:02you've got the risk management and
101. 3:04portfolio manager. The risk management
102. 3:06continuously evaluates portfolio risk by
103. 3:08assessing market volatility, liquidity,
104. 3:10and other risk factors. So they will
105. 3:11adjust the trading strategies on your
106. 3:13behalf, provide assessment reports to
107. 3:15the portfolio manager for a final
108. 3:16decision, and then the portfolio manager
109. 3:18essentially approves or rejects the
110. 3:20transaction proposal after going through
111. 3:22all of these steps. If the trade is
112. 3:24approved then the order will be sent to
113. 3:25the exchange and executed. So to install
114. 3:27this on the computer you want to go into
115. 3:29the GitHub, you want to click copy on
116. 3:31each of these commands and then you want
117. 3:32to put it into Claude code or Codex you
118. 3:35can use either. Or what you can do is
119. 3:36you could take the GitHub link and you
120. 3:38could get Claude to do it itself. But if
121. 3:40you want the most seamless installation
122. 3:41experience possible then in my free
123. 3:43school community down below I'm going to
124. 3:45leave a resource where you can just drag
125. 3:47and drop it into Claude code. It'll
126. 3:49automatically install and it will
127. 3:51automatically be set up with the
128. 3:52framework that I'm showing you in
129. 3:53today's video. So, it's basically a
130. 3:55cheat sheet of sorts that anybody can
131. 3:56start. In my school community, you'll
132. 3:58also find all of the resources and
133. 4:00assets from these videos. So, as you can
134. 4:02see here, if you click on the classroom
135. 4:04and then free asset library, every
136. 4:05single time I upload a video, you can
137. 4:07get the full setup guides and all of the
138. 4:10MD files that you can use to set this
139. 4:11stuff up on your own. And it's good
140. 4:13because everything is in one place, so
141. 4:14nothing ever gets lost. You don't need
142. 4:16to go through a bunch of portals and
143. 4:17claim a bunch of stuff. If you ever want
144. 4:18anything, you just need to go into my
145. 4:20school community and you can access the
146. 4:22resources from all of these videos,
147. 4:24including the resources from today. And
148. 4:26I think that's great because over time
149. 4:27we're building the blocks for you guys
150. 4:29to go from beginner to a seasoned
151. 4:31experienced AI trader. And as new
152. 4:33technology comes out, as new models come
153. 4:35out, as new features and tools come out,
154. 4:37it will only make us better and better
155. 4:39and better. We're only just scratching
156. 4:40the surface of what you can already do.
157. 4:42In 6 months' time, AI trading is going
158. 4:44to be even better. So, I think being a
159. 4:46part of a community that's at the
160. 4:47forefront of that is super valuable. So,
161. 4:49I just put in a prompt to kick off a
162. 4:51research process on Bitcoin. So, I asked
163. 4:53it to run its analysis on Bitcoin for
164. 4:55today, and you are now seeing it's going
165. 4:57through the process of analyzing Bitcoin
166. 5:00in accordance with the initial steps we
167. 5:02discussed before. So, the research team
168. 5:04is curating information that's going to
169. 5:05go through this process. So, I just ran
170. 5:07the entire process on Bitcoin, which
171. 5:09took about 10 minutes, and it put it
172. 5:11through the entire system. So, firstly,
173. 5:13we have the market analysts doing their
174. 5:15research. So, you can see here price
175. 5:16action and trend structure, momentum
176. 5:18MACD, momentum RSI, volatility with a
177. 5:20takeaway for each section. The synthesis
178. 5:23is that Bitcoin is in a confirmed
179. 5:24multi-time frame uptrend that recently
180. 5:27broke out to a closing high of 81K after
181. 5:29a shallow control pullback in early to
182. 5:30mid-September that did not violate
183. 5:32structure. MACD histogram is recovering
184. 5:34sharply from a corrective dip and
185. 5:35nearing a bullish crossover. By the way,
186. 5:37you could even prompt and get graphs to
187. 5:39demonstrate each of these trends. You
188. 5:41can actually format a visual research
189. 5:43report using this information, which
190. 5:45which is really, really cool. You can
191. 5:46even use this information to create
192. 5:47custom indicators. You can ask Claude,
193. 5:49you know, if you connect the TradingView
194. 5:50MCP to overlay this information over the
195. 5:53chart as well. So, that's a really good
196. 5:55way to learn. Price is trading near the
197. 5:56upper Bollinger band, which in a strong
198. 5:58trend is typically a continuation
199. 5:59signal, though it does raise near-term
200. 6:01pullback risk towards the 78,400
201. 6:04middle band. The combination of a fresh
202. 6:06breakout to highs, recovering momentum,
203. 6:08and non-overbought RSIs constructively
204. 6:10bullish, but traders should watch for a
205. 6:12confirmed MACD signal crossover as
206. 6:14validation and whether price can hold
207. 6:16above the 78,400
208. 6:18to 79,800
209. 6:20zone, since that would confirm continued
210. 6:21trend health rather than a failed
211. 6:23breakout. Then you've got the sentiment
212. 6:25analyst who dug through all of the news
213. 6:28and put together its proposal. Then the
214. 6:30fundamentals analyst did its analysis
215. 6:33with the entire summarized bull case,
216. 6:36entire summarized bear case. I mean,
217. 6:38this is really top-level stuff. If you
218. 6:40actually read through the information,
219. 6:41it's very, very in-depth, and I would
220. 6:43imagine it would be particularly useful
221. 6:45when it comes to analyzing things like
222. 6:46stocks, and I'm going to give you an
223. 6:47example of that in a minute here. We
224. 6:48have the trader, we have the risk team,
225. 6:50we have a bullish, bearish, and neutral
226. 6:52analyst giving their opinions. And then
227. 6:54finally, after all of that, it comes
228. 6:56down to the portfolio manager with the
229. 6:58overall rating, which is hold, with the
230. 7:00executive summary saying, "It is not a
231. 7:02place to add and also not a place to
232. 7:04liquidate. Load a conditional upgrade to
233. 7:06overweight on a decisive rally close
234. 7:08above 83,000, with the first tranche
235. 7:10gated on a successful retest of 83K at
236. 7:14support." So, what we can do as the next
237. 7:15step here is actually start to develop
238. 7:17some AI automated triggers, which I'll
239. 7:20show you how to do that to actually
240. 7:21automate a trade. One thing that's very
241. 7:22obvious when you read through this
242. 7:24analysis is that it's not AI slop. Like,
243. 7:26if you ask a Claude chat for
244. 7:27information, it doesn't use the same
245. 7:29language, formatting, depth. It doesn't
246. 7:31have access to the same data sources
247. 7:33that you're obviously getting through
248. 7:35the Trading Agent's resource on GitHub.
249. 7:37So, when it comes to actually analyzing
250. 7:39stocks, Bitcoin, and automated trading,
251. 7:41this is like miles better than if you
252. 7:42were just using a Claude app. We'll get
253. 7:44into the trading in a second, but I find
254. 7:46this particularly useful when it comes
255. 7:48to stocks because stocks have even more
256. 7:50data when it comes to the financial
257. 7:52reporting, projections, fundamental,
258. 7:54bull case, bear case, etc. So, if I do a
259. 7:57prompt like this, could you please run
260. 7:59an analysis from the beginning on Meta
261. 8:01stock to work out if it's a buy or a
262. 8:04sell right now? This is a stock that
263. 8:05fundamentally I'm actually bullish on
264. 8:07because and obviously not financial
265. 8:08advice, but my personal thesis is that
266. 8:11distribution in the AI era will become
267. 8:13really important, especially if
268. 8:15open-source models genuinely threaten
269. 8:17the frontier labs. Then the enterprises
270. 8:19with baked-in distribution, so Google is
271. 8:21one of them due to the amount of people
272. 8:23that use the Google Suite, Microsoft is
273. 8:24one of them due to the amount of people
274. 8:25that use Microsoft, and then Meta is one
275. 8:27of them because of the amount of people
276. 8:28that use Instagram, Facebook, and
277. 8:30WhatsApp. I think given how strong of a
278. 8:32position they're in to distribute AI and
279. 8:34optimize AI in terms of ad spend, I
280. 8:36think they're in a strong position
281. 8:37personally, but I'm interested to see
282. 8:39what the analyst actually says about the
283. 8:41price and whether there's a trade
284. 8:42opportunity. What you can also do is you
285. 8:44can set loops running overnight. So, if
286. 8:46you plug your portfolio into Claude or
287. 8:48into Codex, you can actually have an
288. 8:50analysis happening overnight. So, every
289. 8:52single morning you get a report that
290. 8:53notifies you whether there has been a
291. 8:55material change in your portfolio. The
292. 8:57way to do this is by scheduling a task
293. 9:00on Claude, which you can do. But since
294. 9:01it's on Claude code, you do need the
295. 9:03session to be open overnight. So, either
296. 9:05doing it on a separate device, which is
297. 9:07always on, or on a VPS, so a virtual
298. 9:09private server that you can set up on
299. 9:10one of the websites. For example, I use
300. 9:12Hostinger. If you want a link, I'll
301. 9:13leave it below. I think we have, you
302. 9:15know, a discount going. But I
303. 9:16essentially have a computer running 24/7
304. 9:18for a lot of my AI trading because if
305. 9:20you want truly 24/7 autonomous trade
306. 9:22execution or even, you know, at the very
307. 9:25least analysis, then you're going to
308. 9:26need a computer that's on 24/7. Another
309. 9:29thing is you got to remember this is the
310. 9:31brain because we're interfacing with
311. 9:33Claude code as a harness right now. But
312. 9:35you could develop a front end that looks
313. 9:37a lot nicer than this. For example, you
314. 9:39could use an application like Buzz,
315. 9:41which is built by Jack Dorsey, which is
316. 9:42essentially Slack for agents. You could
317. 9:44have research channels that give you
318. 9:46alerts that are formatted nicer. You
319. 9:47could even create a custom front-end, so
320. 9:50you could build your own front-end UI
321. 9:51dashboard with updates on your
322. 9:53portfolio. So, you just go in the
323. 9:54morning and look at a dashboard and
324. 9:55everything's formatted really nice. So,
325. 9:57this is the raw information cuz I'm
326. 9:59showing you the behind-the-scenes
327. 10:00process of trading agents, but you can
328. 10:02make it look really nice. And of course,
329. 10:03once you start actually executing,
330. 10:05you'll have a nice dashboard because
331. 10:07you'll be able to see your positions.
332. 10:08For example, I had an AI agent place a
333. 10:10position on Nvidia, which I've already
334. 10:12made $40 on, which I'll show you the
335. 10:14execution piece once the analysis is
336. 10:16finished here. So, you can see here that
337. 10:17it's actually breaking down the
338. 10:19year-on-year revenue growth, which is up
339. 10:2028% for Meta, their gross margin, their
340. 10:23forward multiple, the trend, the bear
341. 10:26case. So, I think with stocks it's a
342. 10:28very, very cool tool to use, and you can
343. 10:30actually use this to generate a buy,
344. 10:32hold, or sell rating for your stock
345. 10:34positions, which I think is really
346. 10:36handy. Okay, so there are three levels
347. 10:38to using this tool, in my opinion. Level
348. 10:40one is the morning read. So, you can run
349. 10:42it based on your portfolio. So, you just
350. 10:44copy and paste your tickets in. You get
351. 10:45it to produce a summary, and then, you
352. 10:47know, you can manually then decide if
353. 10:48you want to make any material changes to
354. 10:50your portfolio. It probably costs around
355. 10:52a $1 to $2. It depends on the model, and
356. 10:54it depends on the effort of the model
357. 10:55per run. So, it will cost some money,
358. 10:58but you don't need to run it every day,
359. 10:59especially if you're a higher time frame
360. 11:00trader. Level two, this is the level up
361. 11:02that I was discussing, is it runs itself
362. 11:05based on a schedule. So, you'll need a
363. 11:06VPS for this, where you have a schedule.
364. 11:08So, every day at 9:00 a.m., every 2 days
365. 11:10at 9:00 a.m., you run through your
366. 11:12entire portfolio, which you you can have
367. 11:14your portfolio on Notion or a Google
368. 11:15Sheet, you know, wherever you keep it,
369. 11:16you can just connect it, and you can
370. 11:18have a summary landing in Telegram. You
371. 11:20can connect it with Bot Father. You can
372. 11:21even use Notice or another tool to get
373. 11:23it sent to your WhatsApp. So, you can
374. 11:24just literally get a notification. This
375. 11:26will cost you maybe $30 to $50 a month.
376. 11:28It depends on how often you want the
377. 11:30alerts, but that is for the pure
378. 11:31analysis side of things. So, those are
379. 11:33the first two levels. The third level,
380. 11:35which I'll show you now, is essentially
381. 11:36putting these trades onto a real
382. 11:40execution book. I'll show you a paper
383. 11:42trading book in a second, and then you
384. 11:44can level it up to, you know, a real
385. 11:45money account if you want. The principle
386. 11:46is the exact same, where you get the
387. 11:48outcome that is then translated into a
388. 11:50plan, and then that plan, which is
389. 11:52developed by the engine, is then treated
390. 11:53as a conditional order which gets placed
391. 11:55on a brokerage, for example, Alpaca. And
392. 11:57obviously, you can have an MD file with
393. 11:59a set of guardrails, for example, max
394. 12:0120% per position, no market orders, so
395. 12:03limit orders only, cancel stale orders.
396. 12:05You can have a kill switch file. You can
397. 12:06even have an approval, where just before
398. 12:08it sets the order, you need to manually
399. 12:10click approve. So, it will line
400. 12:11everything up, the buy price, the
401. 12:13amount, etc., and you just need to
402. 12:14approve and click approve. That's one
403. 12:15way that you can reduce, because I think
404. 12:17a big thing, you know, with crypto
405. 12:18trading, stock trading, you know, people
406. 12:20have complacency or they don't act. Um
407. 12:22I've been the exact same for a long
408. 12:23time. Like, I'll have an idea, I'll have
409. 12:24a trade thesis, but when push comes to
410. 12:26shove, I don't act on a position, buying
411. 12:28or selling. Whereas, this takes away the
412. 12:30friction and literally makes it a binary
413. 12:32yes or no. So, it takes away all of that
414. 12:33grunt work. But, I'll show you right now
415. 12:34how to actually conduct level three. So,
416. 12:36we just analyzed Bitcoin and Meta. So,
417. 12:38I'm going to ask Claude to read those
418. 12:40two reports, Bitcoin, Meta, turn each
419. 12:42verdict into a concrete order plan for
420. 12:45my Alpaca account using paper for the
421. 12:47example of this video, but in the top
422. 12:48left you can switch to a real account.
423. 12:50But, I always recommend when you start
424. 12:51learning automation, when you start
425. 12:53testing, always test using a paper
426. 12:54account, and then, you know, build
427. 12:56systems which are robust, and then you
428. 12:57can always move to real trading later.
429. 12:58Same with algo trading bots. Always
430. 13:00start on a paper exchange, and then move
431. 13:02to real trading later. But, the process
432. 13:04is the exact same. I've given it some
433. 13:05rules, so each order, maximum 10% of
434. 13:07current cash, no market orders,
435. 13:09everything's a limit or a stop limit
436. 13:10that only fills if the desk's trigger is
437. 13:13hit. If the verdict is hold with a
438. 13:14breakout trigger and a pullback zone,
439. 13:15place one order for each, a stop limit
440. 13:17buy at the breakout level, and a limit
441. 13:19buy in the lower half of the pullback
442. 13:21zone. Before placing anything, show me a
443. 13:22table, tick a verdict, order type,
444. 13:24trigger price, limit price, quantity,
445. 13:26dollar size, etc. Wait for me to type
446. 13:28yes to process the order. After I
447. 13:29confirm, place the orders through the
448. 13:31Alpaca MCP, then show me the accounts
449. 13:33opening order so I can see them on my
450. 13:34book. So, you can see here it took my
451. 13:36instructions and it developed a real
452. 13:38trading plan. So, you can see here we've
453. 13:40got the stop limit buy based on the
454. 13:43analysis at $83,000.
455. 13:45So, that would be the breakout trigger
456. 13:47zone. We have the meta stop limit buy at
457. 13:49$703. We have the meta limit buy at 620
458. 13:52as well. It gives us the justification
459. 13:55for the range and it also points out the
460. 13:57nuance so we can set a limit order, but
461. 13:58if it's not a decisive daily close, then
462. 14:01we might cancel the order last minute.
463. 14:02That's where the manual approval comes
464. 14:04in. And of course, the orders would have
465. 14:06to be conditional on a macro check at
466. 14:08the time. You could theoretically also
467. 14:09run an experiment where you just gave
468. 14:12the trading agents full range of your
469. 14:14account. And you know, if you use paper
470. 14:16trading for example, you can literally
471. 14:17get them to make adjustments whenever
472. 14:19they want on your positions or a
473. 14:22specific set of assets that you asked
474. 14:23them to analyze. But in practice, you
475. 14:25probably have a portfolio that you're
476. 14:26managing. You probably have assets that
477. 14:28you're interested in and you're probably
478. 14:29still wanting to have some sort of
479. 14:32manual discretion over the the entries
480. 14:34and exits on those assets. So, a system
481. 14:36like this in that case would make more
482. 14:37sense. So, you can see it's just placed
483. 14:40all of the orders. So, if we go onto my
484. 14:41exchange account, you can see we have
485. 14:43the meta limit order. We have the
486. 14:44Bitcoin limit order and uh these were
487. 14:46previous buys that I made on another
488. 14:48test. And we also have an expiry date in
489. 14:49case the orders don't trigger by those
490. 14:51dates. It's quite a cool feeling being
491. 14:53able to talk to Claude and then having a
492. 14:55brokerage trading on your behalf. As I
493. 14:57said before, if you run a VPS, you could
494. 14:58literally run an experiment where you
495. 15:00just get a model like this to trade on
496. 15:02your behalf. I probably wouldn't
497. 15:03recommend that though because I think
498. 15:05the power of this model is empowering
499. 15:07you with the tools to make better
500. 15:10decisions and to speed up execution. And
501. 15:12I think the concept of speeding up
502. 15:14execution is so important because we all
503. 15:16have busy lives. So, if you can just get
504. 15:17an update on Telegram, be on the way to
505. 15:19work on your phone, and then just send
506. 15:22it a message like, "Yep, trigger the
507. 15:23position, execute the stop loss." And
508. 15:25just have that executing. That's so much
509. 15:26better than needing to sit down, log
510. 15:28into a brokerage account, manually amend
511. 15:30a position, set a limit order, check
512. 15:32that it's correct. I think the future of
513. 15:33trading is people interacting with
514. 15:35agents directly. So, I'm going to leave
515. 15:37the full setup guide from this video so
516. 15:38you can just drag and drop it into
517. 15:40Claude. It'll include all of the steps.
518. 15:42It'll walk you through the steps on
519. 15:43setting up also an Alpaca account. You
520. 15:45can also use other MCPs for other
521. 15:47brokerages, so you don't need to use
522. 15:48Alpaca. I just find it's the best and
523. 15:50the most compatible with the majority of
524. 15:52the AI models. And in the free school
525. 15:54community, we'll have all of the other
526. 15:55resources from all of my other videos as
527. 15:57well. So, let me know what you guys
528. 15:58think of trading agents. I think it's
529. 16:00one of the best AI trading free GitHub
530. 16:03repos out there. Subscribe to the
531. 16:05channel for more content like this, and
532. 16:06I will see you in the next video. Have a
533. 16:08lovely rest of your day. Peace out.