I Built an AI-Powered Crypto Trading System in One Afternoon

There is a gap between "AI can trade crypto" and actually wiring up something that works. I spent a Saturday afternoon closing that gap: one Claude session, four data services, and by dinner a system that screens tokens, sizes positions, sets stop-losses, and places real trades on a paper account. No code. No API wrappers.

Here is what I built, what surprised me, and why I think it matters for anyone watching where AI-assisted investing is heading.

Update — June 2026: This post describes the v1 system as I built it in April 2026. The strategy has since been rebuilt around a single load-bearing rule — a 200-day macro trend gate — and the LunarCrush-based sentiment strategies described below were retired when LunarCrush discontinued its free API tier. This write-up stands as a record of the starting point. For what's running today, read the v2 rebuild and the stress test, or watch it live on the dashboard.

Four data sources, one brain

The idea was simple: most trading mistakes come from looking at only one kind of information. What if an AI could hold four kinds at once?

I connected four MCP servers to Claude. (MCP — Model Context Protocol — is a universal plug that lets an AI assistant talk to outside services.) Each one answers a different question:

• Alpaca is the hands. It places orders and manages stop-losses on a $100K paper-money account.
• LunarCrush is the crowd. Real-time social sentiment for every major token — the why behind a move.
• Crypto.com is the tape. Price history, order books, volume — the when.
• Blockscout is the chain. Who actually holds a token and where it's flowing — the who.

None of these are new on their own. The interesting part is what happens when one brain reads all four at the same time.

What the screen turned up

I asked Claude to run all four sources across nine tokens. What came back was not a "popular coins" list — it was the kind of cross-checked picture that used to take an analyst days.

Solana had the highest social sentiment of the group, driven by real developer news — but the same data flagged a flood of Solana memecoin scam content on TikTok. A mixed signal that a price chart alone would never show.

Chainlink looked strong on sentiment, but the blockchain told a second story: Chainlink Labs controls roughly 210 million LINK across identifiable treasury wallets — a sell-pressure overhang the social feeds were already starting to whisper about. Strong story, capped position size.

Two tokens got cut for the most boring reason possible: trading volume on the paper exchange was too thin to fill orders. I learned that one the hard way when an earlier order sat unfilled for hours.

Putting $20K to work

Based on the screen, I deployed $20K of the $100K account across six positions — heaviest in BTC and ETH, a conviction position in SOL, smaller stakes in AVAX, DOGE, and LINK. Every position got a server-side stop-loss: tighter on the blue chips, wider on the speculative names. Those stops live on the exchange, not on my laptop — crypto trades 24/7 and I don't.

Selling is the hard part

Buying is easy. The real work was building exits that don't depend on me being awake. I wrote seven sell rules that run automatically three times a day. A few examples:

• Sentiment collapse: if the crowd turns on a holding, sell before the price fully reflects it.
• Trailing profit lock: once a position is up 15%, protect the gain — pull back too far from the peak and it sells automatically.
• Whale alert: if those Chainlink treasury wallets start moving tokens toward exchanges, that's a sell signal days before the market feels it.

Each rule watches a different data source. Sentiment collapse comes from social data, breakdowns from price data, whale alerts from the blockchain itself. No single source tells the whole story — the edge is the combination.

The full setup — MCP configuration, every wallet address, all seven rules — is in the companion technical guide.

What surprised me

Rate limits are the real constraint. The free social-data tier allows so few calls that one full portfolio scan nearly exhausts a day's quota. Clever business model: the AI ends up recommending the upgrade because it literally cannot finish the analysis without it.

On-chain data is genuinely differentiated. Watching whether specific treasury wallets are moving tokens used to require a Bloomberg terminal and an analyst. Now it's free and programmable.

The plumbing has opinions. Alpaca won't let you buy a token while a stop-loss sell exists on it — you have to cancel the stop, buy, then recreate it. A small detail that would silently break an automated system that didn't know.

Synthesis is the actual product. Every individual data source is getting easier to access. The value is in combining them into one coherent decision — which is exactly what an AI holding all four contexts can do.

What this means

The tools to build a multi-source trading system are now available to anyone who can describe what they want in plain English. A year ago this required a quant team. Six months ago, a developer. It took me a Saturday.

This is a paper account, not real money. But the system is live, and every position, trade, and strategy trigger is tracked in real time on the AI Crypto Strategy Dashboard. If you're experimenting with anything similar — different data sources, different rules, edge cases I haven't hit — I'd love to hear about it in the comments.