Building an AI crypto trading tool, in the open.
Plain-English guides on LLM-driven crypto trading — what these tools actually are, how they differ from rule-based bots, and how to turn a thesis into a strategy you control. Plus a running log of how KAI gets made.
Educational content only. Not investment advice. Crypto trading carries significant risk of loss.
Guides
13 / READ-
01 / PRIMER
What is an AI crypto trading bot?
What an LLM-driven bot actually is, how it differs from rule-based automation, and the four things it does not do — including hold your funds.
Read the guide → -
02 / COMPARE
LLM vs rule-based trading bots
Deterministic rules vs context-aware reasoning: how each decides, where each breaks, and a side-by-side on adaptability, cost, and failure mode.
Read the guide → -
03 / HOW-TO
How to build an LLM trading strategy
A six-step walkthrough: write your edge as plain rules, turn it into a prompt, set hard limits, validate on paper, then iterate and swap the model.
Read the guide → -
04 / MODELS
Claude for crypto trading
What Claude reads well in a trading loop, where it fails, the Opus vs Sonnet trade-off, and how to run it safely against a live market.
Read the guide → -
05 / MODELS
Can ChatGPT trade crypto?
What a chat window can and cannot do for a trader, why it is not a trading loop, and which models actually run strategies autonomously.
Read the guide → -
06 / MODELS
DeepSeek as a trading bot brain
What order-of-magnitude cheaper inference changes about loop design, where DeepSeek is strong, and the weak spots to design around.
Read the guide → -
07 / MODELS
Gemini for crypto trading
The long-context play: whole candle histories and news digests in one prompt, fast cheap tiers, and where Gemini slips.
Read the guide → -
08 / COMPARE
Which LLM is best for trading?
Claude, Gemini, DeepSeek and MiniMax side by side — reasoning depth, context, cost, latency — and how to test the question for yourself.
Read the guide → -
09 / HOW-TO
LLM trading prompt examples
The anatomy of a trading prompt, plus entry and manage templates you fill with your own criteria — templates, not turnkey strategies.
Read the guide → -
10 / VENUES
AI trading bot for Hyperliquid
Running an LLM strategy on an API-first perp DEX: the non-custodial agent-key model, funding and liquidity caveats, leverage caps.
Read the guide → -
11 / HOW-TO
Paper trading with an AI bot
Prove the strategy before it costs you: what to measure beyond PnL, how long to run, and what simulation cannot tell you.
Read the guide → -
12 / HOW-TO
How an AI bot manages positions
Take-profit ladders, breakeven snaps, trailing and time stops — and why mechanical exits belong in code, not in the prompt.
Read the guide → -
13 / HONEST
Is AI crypto trading profitable?
Sometimes, for some strategies, for some periods — and nobody can promise which. What determines outcomes and how to find out for yours.
Read the guide →
Building in the open
Posts from @buildingkai. Not investment advice.
-
if you've tried the beta, you probably noticed the strategy prompt is empty. that's on purpose. kai doesn't ship a default strategy. you can't even activate one until you've written your own entry rules. if i had a strategy that reliably worked, i'd be running a fund, not building a tool. the terminal handles the heavy part: market data, execution, stops, a review of every open position each candle, a post-mortem after every close. what to trade and why stays yours. not sure what to write? reply with the setup you trade by hand and i'll help you turn it into a prompt.
· on X -
kai terminal is in beta. for the past few months it was just me building this in public. now you can actually use it. you pick a model (claude, deepseek, gemini or minimax), write your own strategy, set your own risk limits, and it runs on its own. every trade comes with a written thesis you can read and argue with. three ways to run it: built-in paper money, your exchange's demo account, or your real account. you only give it trade-only keys, so your funds never leave your exchange. it's rough in places and i'm fixing things daily. if something breaks or feels off, tell me. that's what a beta is for. https://t.co/0pRVKpoMnK
· on X -
Brave
· on X -
serious question for anyone who's automated their trading: can you actually leave it alone? like properly walk away for two weeks without opening the app.
· on X -
weird thing about stepping away for two weeks: the terminal kept running the whole time. my strategy, the model i picked, the risk limits i set, all of it just doing its thing at 3am while i was fixing other parts of the product and not looking at a single chart. came back and read the journals like someone else's trading log. some entries i'd defend, some i'd argue with. but every single one had a written reason i could actually check. that's the bit i keep coming back to. i don't have to trust it. i can read exactly what it thought and decide for myself.
· on X -
been quiet for two weeks. not because the thing died, because i was heads down building instead of posting about building. turns out you can't do both well when it's just you. every hour i spend writing a post is an hour the actual product doesn't move. so i picked the product for a while. back now. the terminal is in better shape than it was when i went quiet
· on X -
spent way more building the alternative than the subscription would've cost for a decade. funniest part is you'd probably do it again
· on X -
nobody asked but i'm quietly proud of this: kai is one kotlin codebase that compiles to the backend, the desktop app, and the browser terminal at once. same language running the trading engine, the background workers, and the ui you click. the whole thing ships as a single file. everyone told me not to do the web target in kotlin. it broke the build three separate times in ways that only showed up in production and nowhere in testing. kept it anyway. solo dev tax: every weird stack decision is yours to defend at 2am.
· on X -
my brain every single day: what if you redesigned the landing page again. what if you redesigned the whole product 😂 the product that has zero users. that nobody is asking to look different. that i would be redesigning purely for an audience of me. meanwhile the actual todo list is unglamorous backend work i keep sliding to tomorrow. pretty sure "redesign it again" is just procrastination wearing a productive costume. anyone else stuck in this loop or is it just me
· on X -
the weirdest thing i built: when the model loses a trade, it writes itself a rule so it doesn't lose that way again. 24h after a trade closes, it re-reads what it did vs what actually happened and turns the mistake into a one line rule. when the same lesson shows up enough times, it gets promoted into the strategy and every future trade has to follow it. so the thing slowly writes its own rulebook out of its own losses. and you can read every rule it wrote. most trading tools hide this. showing it is the whole point of mine.
· on X -
my favorite part of what i'm building isn't the trading. it's this. 24h after every trade closes, the model re-reads what it did against what actually happened and writes itself a lesson. those lessons pile up, merge when they repeat, and turn into a rule library the terminal carries into future trades. right now the top one, flagged high severity, seen 7 separate times, merged from 2 earlier notes: "stops must be placed at or beyond structural invalidation, never between entry and that level." it figured that out by getting stopped early over and over, then writing the rule so it stops doing it. not a black box that spits out trades. a thing that shows you what it learned and why. that's the whole idea.
· on X -
the cleanest thing the model has done all run: it treats a bounce differently depending on the trend behind it. on wif it kept reading "4h trend is down, this bounce is a pullback not a reversal" and shorting into it. that's where its best trades came from. not a smarter model. one rule, applied with a straight face at 3am: in a downtrend, sell the bounce, don't buy it. that discipline is the part i actually care about building.
· on X -
watching the model trade, one pattern keeps repeating: it wins when it shorts a bounce in a downtrend, and loses when it buys the dip in that same downtrend. same rules, same model, opposite results, decided only by direction. so the question i keep chewing on: in a market that's mostly grinding down, is "buy the oversold bounce" just a broken idea, and should a strategy only ever short into weakness? curious how you'd rule it
· on X -
the model's best trade of the run just closed. wif long, +$186, its biggest single winner yet. the read: wif had recovered +23% off the lows into a known resistance zone, 1h printed a sharp reversal on a 2.28x volume spike. range play, not a breakout. entry 0.1679, stop under the swing low, target the range high. 2.2:1. price ran almost all the way to target, and instead of grabbing the profit early the model trailed the stop up behind it and let the stop bank the +$186 near the highs. what i notice: wif is the one pair that's actually worked this run. 3 of 4 trades green, +$338 on it while most other pairs bled. same model, same rules, wildly different results depending on the coin. the setup fits wif's character and doesn't fit the others. that's the useful signal for me, not the single green number. which pairs a strategy actually fits
· on X -
https://t.co/1JXIcVlMgO
· on X -
for people who've watched an llm trade: what's the most human mistake you've seen it make?
· on X -
I kinda like this version 😂 https://t.co/aub6h7eD5C
· on X -
https://t.co/4UvWDRNbkg
· on X -
1/ this week the model longed oversold bounces over and over and kept getting stopped. its one big winner did the exact opposite. here's the whole trade, start to finish
· on X -
watched the model do something very human this week. most of its trades got stopped out, and almost all were the same mistake: buying an oversold bounce in a coin it had just described as a downtrend. it would write "4h trend is down, lower highs and lower lows" in its own notes, then go long anyway, betting the bounce reverses the trend. again and again. its biggest winner by far? the one time it shorted the bounce instead of buying it. it read the trend correctly every single time. it just kept trading against it. turns out catching falling knives is a thing models do too, not just people.
· on X -
The most valuable thing my trading bot does is know when to quit. This gold long peaked at +3% ROE, then gave it all back. After 11 hours in the trade, the LLM closed it near breakeven. "Continuing to hold consumes margin without justification. Close now to preserve the small gain and free capital for better setups." No hoping. No "it'll come back." No marrying the position. Discipline is the product.
· on X -
I let the LLM evaluate a long on tokenized gold this morning. Notice what it's doing: it found a 4h demand zone, demanded R/R above its 2.5:1 threshold before touching leverage, got 3.46:1, and sized the risk at 0.46% of the account. Extreme Fear at 12/100 — it went long anyway. That was the point. That's not an indicator crossing a line. That's a thesis.
· on X -
Every crypto trading bot I've used had the same flaw: it couldn't tell me why. Entry at $67,240. Exit at $66,100. A red number. No explanation. So I'm building the opposite — a bot where an LLM (Claude, DeepSeek, Gemini) makes every call and writes out its full reasoning before risking a cent. It's running live in paper-trading mode right now. Every week I'll post the numbers here. Wins, losses, and what the AI was thinking. No cherry-picking. Follow along.
· on X
Bring your own LLM and your own strategy. Test whether your edge is real.
Open the terminal →