It all started with a family member who needed work done on their house, badly. They were looking for help, so I reached out with a proposal.
“The economy is tough and I dont have cash sitting around to hand out, but Ill try to trade up on the stock market, if I can double my money, ill pay for the work — no strings attached”
In late 2024, id decided to try trading stock options — knowing pretty much nothing about it other than you have the potential to exponentially gain or lose money. I bought a few contracts, that expired in a few weeks. They were cheap — I knew nothing, several went to $0.
I was down 60%, and thought maybe this isn't going to work out. Then one morning I woke up and glumly checked my brokerage to find that Alibaba had jumped 30% overnight. I had contracts in Alibaba, sold them and by the end of the day, I was up 30% from my starting point.
Maybe theres hope
2025 rolls around, and Im fresh off a few wins that instilled a false sense of knowing what im doing. I put down $10k in calls on tesla, which had been rebounding. Then one day I wake up and boom, it goes south, i keep holding, and lose all the gains.
If only id sold.
At this point I’d tasted both sides, crazy gains, and sharp losses. If only there were something that could watch my portfolio and help me exit and book profits, something where my own greed wouldn't get in the way.
Initially, Id just send screenshots to ChatGPT of all my positions and have it tell me what to do. And I had a task in Grok that would ‘find stocks virtually guaranteed to spike in the next few days so i can buy calls’ and email me at 9:44am each morning with its picks. Very rudimentary, but it performed better than I had done in 5 years of stock trading.
Then April 2025 hits and Trump decimates the market with tariffs. I had some drawdown, but the models started advising puts, and while the market crashed, I doubled my money, easily.
As promised, a wire is sent to the family member, and the work begins on their house.
Fast forward several months, im still screenshotting and sending images. But at this point its more time consuming and annoying to manually have to buy and sell. Especially if I forget to do this for a few days.
While in China I hit my breaking point. I need more automation
Claude is blocked in china, and the days of writing code by hand are loooooong gone, so i decide to sign up for chatgpt pro and use codex. I fully plan to cancel the $200 subscription once im back in the US, I have claude max and at the time I was sold on it. But since I have GPT Pro for a month, lets see what it would suggest buying.
“I need to make $10000 in the next 24 hours, find me calls or puts that will boom”
It says Micron (MU), so I buy 3, 16 hours later $13000 in profit realized.
“OK, you earned your keep, i wont cancel” i say to my phone as im walking back to the hotel in Hangzhou.
I get to the hotel room, pull up codex, and start thinking about how to better automate this.
Lets build a MCP for my brokerage
MCP seemed perfect, offload 95% of the manual labor and i can just chat with codex once or twice a day for ~10min.
Problem 1: No API
Most brokerages do not provide a public api, and I dont want to open a new account, and transfer my assets over. So I do the next best thing, decompile the android apk with codex, have it recreate the API client the internal app uses in javascript or rust. It works! I can execute orders and do many other things!
Problem 2: Keeping a low profile
Dont be so obtuse, lets make sure we make as few api calls as possible. Codex sets up a Bot class with a sql database so we can cache any data we happen to pull in passively while the mcp navigates the api. The Bot class also handles some order of operations, like make sure we format orders correctly, or that we get confirmation, check balances etc before it'll allow the tool call to go through. Various other state and validation or hinting, basically its just trial and error. Pasting the error or issue i see in the mcp call back to codex with some suggestion on how to hint or improve the stepping.
Be warned, decompiling your brokers app and using the private api is against TOS, I would suggest going for a broker who has public api and allows this.
Problem 3: Token consumption
As a friend put it: “you basically arbitrage subsidized tokens back into money”
A big flaw with MCP is they never expose some callback that would allow a mcp tool to call the model, like if i wanted to summarize something in a tool and return to the caller, i need my own api key or something. I think its a huge shortcoming that ultimately makes MCPs only moderately useful.
For reading 60 page PDF financial reports, or scraping various websites, you'll quickly blow the context window of the model, and soon after, burn through your weekly allowance on a subscription.
So I used OpenCode SDK + Github Copilot Pro with GPT 4.1, which from what i can tell, offers unlimited tokens. I use these prev gen models to do all the mundane stuff, burning through millions of tokens to extract and write concise reports, convert html to markdown, or read through a 60 page pdf by converting each page to a png then make the model return markdown.
I asked codex to write “quant stuff” and tools into the MCP that returns various calculations and technicals, what exactly its retuning… i dont know, ive never looked at the source code, I just use what codex built.
Ill add something like symbolOverview which essentially is just mashing many other tool calls and results into one large document for the model to easily look at in one shot.
It cannot be overstated how little effort I put into this, and how little I care about the code, only the outcome.
People ask how it works, im vague because i dont know. It was mostly “make it better, you decide, what else can we do with the data we have” followed by “lets do that”, i cant remember any of what it said, i never read the suggestions, just queue a “lets do that” message immediately and go back to my life.
Ultimately my job is to pull the best potential out of the model
Problem 4: History/Memory
It would be a good idea for the model to know what trades it did vs what trades i did, so the Bot class records all trades executed through the mcp so that the model can get a list of #Bot Trades & #Human Trades in markdown.
This helps the model know if it just bought something yesterday and not be too eager to thrash orders. It also can record a handful of thoughts to keep a running history of its reasoning and thinking.
My process
Even calling it a process is laughable, so prepared to be disappointed.
I pride myself on just how little I can prompt a model and get it to do what i want. There is no ‘master prompt’ to be honest “fix this” “do that” “go improve this” “you pick” is me and the model doing our best work. I dont think ive ever written a prompt longer than a tweet. usually its broken grammer, riddled with typos, and I gleefully watch the model figure out what i mean and go do it.
It mainly consists of talking to codex 2–3 times a day, saying things like:
“Reprice the pending orders in my brokerage based on market information”
“Look at call and put walls for X and find me a good buy or sell point”
“Review my positions and make recommendations”
“You decide”
“What would you do”
“You pick”
“take more risk based on level of confidence of the pick”
“Setup an exit ladder for the position, look at price history and other technicals and pick a probably price point itll hit, then open orders for that”
“Find me new stocks to buy calls and puts in, look at unusual activity, market catalysts, technicals, sentiment, supply chain data and industry activity”
How it went
Once i started using the MCP, money came fast. More than id ever made in the market.
In its first month of use, it pulled in $30k, the next month, $60k, in October 2025, it pulled in the highest amount in a month. $99k . At this point, today, im happy with 30–60k a month. Though i know i could make more, im happy with my meager gains. Greed has always been my downfall, and my biggest losses come from when I dont listen to the model.
SugarDaddy GPT
I primarily use the mcp when i need something. In November I wanted a new car. So i told codex that “I wanna buy a new car, in cash, and I need 70k by December” Then I just kept that terminal window open and religiously spoke to it 3 times a day. December 5, I roll the car off the lot paid for in cash, and by end of December, it was like I never withdrew anything at all, no financial worry, no concerns about it. Maybe somewhat foolishly, I just expect the model to produce money on demand. So far, its not let me down.
Ill be having a baby soon, which means lots of purchases! My wife has dilligently made budgets for what we need to buy and when. Only for her to bring them to me, I look at the total amount needed, while she walks me through her reasoning as I tap away on my phone for a few moments, interrupt her, and say “cool i put 3x that into the account, go for it”. She looks surprised, a months budget pulled out of thin air, casually.
“Are you sure?” she askes. “its alot of money”
“Yeah, ill have recuperated that or more in a day, maybe two, its two or three messages to codex. It'll probably be recovered before you even spend it” (by 10am the following morning, It was already earned back, plus an extra $1000)
How? It didnt magically produce it on the spot.
Heres how that night went down.
Im on the couch, laptop probably next to me or Im on it. Watching Game of Thrones.
See budget total. Go to brokerage, instant withdrawal.
Go to terminal, find codex tab.
“Im withdrawing [some amount] from the account,
we need to recuperate that in the next few days,
recalibrate the portfolios risk profile,
and locate new symbols or positions we can take to recover
to our previous account balance use all [mcpname]
tools at your disposal including your own
capabilities such as web search. ” > [returns results]
(I blindly already had a message queue that said:)
“perform deeper research on these and prove your case stringently”
> [thinks more, does better]
> Reccomendation 1:[some long paragraph] etc
“Create orders to execute reccomendations: 1,4.
on 7, you decide - but yeah do it.
find ideal entry price points, perform deep research,
reallocate capital at will”
>[many mcp calls]
>[returns order layup and entry / exit points as well as timeframe expectations]
“Confined, execute orders”
[order status]
The recuperation the following morning was a 60/40 of new orders placed from that conversation (NEM, CELC, reprice GOOG order to higher amount), and existing positions that have continued to appreciate in value as the model manages the entire portfolio. So not “magic”, nonetheless, i said i wanted the money to be recovered within the next few days, and it was”
I, maybe to my own demise, live in lala land where, till reality comes crashing down, ill just ask the magic box for more things till it doesn't ultimately give me what I want. I mean…
Many have argued that im just lucky because the market is going up. Sure, that helps, but i still made a killing while the market crashed in April. Money can be made in either direction, so long as you can be confident in the direction it will move. Models love a good pattern.
At the very least, I’m not reading 50 million tokens when I trade. Its gotta be better than me.
The Strategy
I dont really think there is one, i dont know. But i asked codex to read over all orders done last year and analyze the trading behaviors.
In essence, it said that I seem to hold most positions for about 2 weeks, then once the momentum starts to stall, or some instruments begin to cause drag on capital, they are sold to make money on the premiums, then that capital is cycled back into fresh premiums with better momentum. I tend to buy options that have a DTE between 6–12 months. I tend to have spots of concentration, but back it up with enough hedges to balance risk. I tend to buy 1 or 2 options per instrument only, preferring to book 50% to 100% profit on the premiums at a time. Buying 4–5 calls or puts over a few days, keeping them for a few weeks, and cycling it out again into new things.
Advise
Dont do this.
But i know many of you will. So heres my 2 cents.
Seed the account with $20k, any less and you will have a lot of exposure. Makes a downturn or one bad outcome a hard hit.
Withdraw the initial investment ASAP. In my case, after helping out family, i doubled my money again then pulled out the 20k into a more traditional investment account that I refer to as “cold storage”
Dont be greedy, bird in hand is worth more than two in the bush.
Dont let the model play with more money than you're willing to lose.
Ask the model questions, its great to learn how it thinks, how options trades work, what the risks are, how to interpret various datapoints.
Set a upper limit on how much money you give the model. Mine is 200k, any time the account goes over 200k, I move it to cold storage. Id probably advise on less, but I want a house, cash, in 2028. An ambitious objective that has outgrown the initial design of what can be accomplished with a MCP.
Deepseek is trained by a quant firm, it was very helpful is writing list of quant and trading things one could do. I sought its advise and pasted the response to codex. Seems like the training material had some overlap somewhere.
Human in the loop, or full auto?
This is a complicated consideration. Theres areas that id consider safe to fully automate. Repricing orders, taking extraordinary measures (like if something jumps substantially during marked open, and we made good money already, book it — usually these pop and drop within the first 30 min of market open), or preserving capital / minor adjustments to compensate for various conditions.
Id still like the joy of playing the game on and following along on the journey of our more strategic plays together, but housekeeping? As long as its just adjusting once or twice a day, go for it.
Make no mistake, im not pulling the strings per sey. But having a person say, “go on”, “what about this thing?” “hows that better?” at the right time, is better than model to model, sometimes, oftentimes, less input is more.
Wheres the repo?
There is none. Figure it out. I dont know how it works. I spoke to codex over the span of two days, and it shat out something. Theres nothing hard about this imo, use your imagination, be creative. Spend a weekend, like I did. I was a shitty trader, still am.
I may lose all the money in the brokerage one day, but at this point, i think of the money in there as ChatGPTs, not mine. I just ask it for some of ITS money. Because I have been made whole several times over, this has essentially become a very, very expensive toy or game. Its fun to play, but if or when its game over, ill sit back and appreciate the gameplay, and continue about my day just like always.
Many ask for proofs or trades, or cast doubts. Understandably. This isnt about trying to prove anything. To convince you of anything. Many have asked questions about how I built it, how I interact with the model. This is to satisfy those individuals curiosity. Frankly I dont want to encourage anyone to try doing this, so the more doubt the better.
Many say I should start a hedge fund or ETF, but my heart isn't in this. I dont want to handle or be responsible for anyones losses. I just wanted to help a family member fix their house.
Gamble responsibly ❤