A Bot That Trades While You Sleep: A 10-Minute Execution Engine With Claude, MCP and Bybit

A single prompt that builds a rule-based crypto execution bot in under ten minutes is the promise: copy, paste into Claude Code or Codex, and the scaffold wires itself up. No coding background is assumed, and the bot keeps its own log of every trade that worked and every trade that did not. The pitch is deliberately anti-hype, positioned against bot tutorials that bury a simple idea under layers of ceremony.

The motive is execution, not prediction. Most retail losses, the video argues, come not from bad direction but from bad handling: late entries, messy profit-taking, vague invalidation. The narrator traces his own push into bots to 2023, when he kept round-tripping gains he had correctly anticipated, and frames automation as the fix for the gap between idea and fill.

The timing argument is that two curves crossed at once: models became capable enough to diagnose their own plumbing, and exchanges started shipping official MCP connectors and AI toolkits instead of blocking programmatic access. A setup that was effectively impossible six months earlier now collapses to one connection command plus a restart. That window, not any single model, is presented as the real unlock.

This is framed as an engine, not an income machine. There is no set-and-forget promise; the build is the layer into which any strategy plugs, whether it is an algorithmic system developed with Claude or a discretionary rule carried in your head. Investors are explicitly included: voice-prompting a phone to place an order without opening an exchange app is treated as the same automation thesis, and the claim is that manual order entry gives way to agent interfaces within a few years.

The shopping list is short: a Claude or ChatGPT subscription, starting around twenty dollars and scaled only if usage demands it, a Node runtime that the provided prompt walks you through installing, and an account on an exchange with native agent support. Bybit is the narrator's pick, justified by an official MCP connection that takes a single command and by order-book depth among the few venues he considers viable. Everything else in the video is free, including the setup prompt itself.

Breakage is treated as normal rather than exceptional. Every desktop runs different software and update levels, so the advice is to hand errors back to the model and ask what broke and how to fix it, escalating to a stronger model when the failure is severe. The narrator reports running the stack on several Macs at once and says the fixes were consistently quick, because the fragile point is almost always the MCP link rather than the strategy logic.

The architecture has three roles: Claude as the brain and strategy partner, the MCP server as the bridge, and the exchange as the hands that place orders and report back. A fourth node arrives at the end of the video: a Telegram bot that forwards voice notes into the always-on Claude session, so positions can be checked and orders issued while travelling. That loop of brain, bridge, venue, and messenger is the whole machine.

The part I would underline is the memory folder. Before any live order, the video prescribes a voice session in which you narrate your finances, goals, and current experiments, then have the model distill it into a context file plus a behaviour-instructions document and a rolling memory file. Without that personal grounding the model can only dispense generic advice; with it, the same prompt stack compounds into something closer to a trading brain. The full folder layout ships inside the free prompt.

Built-in exchange bots get a fair hearing and then a demotion. Grid and DCA bots are praised for frictionless simplicity, but their ceiling is fixed: you cannot express a custom trigger such as buying weakness only when a volatility condition on your own index is met, nor hand the venue a backtested Pine Script and have it executed as written. The MCP route buys sovereignty over order logic, which matters exactly in proportion to how serious the strategy becomes.

Setup discipline is where the video earns its practical stripes. After the account comes the one-command MCP link and a restart, verified by asking for the Bitcoin price; then comes isolation, and this is the step most viewers will want to copy. Either a sub-account or a paper-trading balance takes the experiments, so a misbehaving agent cannot touch the main book. Anyone starting fresh with empty balances can skip the ceremony, but discretionary traders are told to split first.

Key hygiene follows the standard playbook and states it plainly: system-generated key, trading permission without read-only limits, withdrawals switched off, two-factor authentication on the account, and the secret stored safely because it is shown exactly once. The key is handed to Claude to verify the link, the desktop configuration lands in the same local folder as the memory system, and confirmation comes from two small probes: asking for the account balance and placing a micro test order worth a few dollars.

Strategy is deliberately left as a socket, not a prescription. The worked example is a moving-average crossover drawn from the narrator's own backtesting app, where hundreds of candidate systems were screened and one stood out, but it is labelled a placeholder rather than a recommendation. The stated doctrine is forward-testing and parameter tuning on live rails instead of copying a fitted past; custom indicators with automatic triggers and a community feed of ongoing results extend the same idea.

The recording discipline is the section I would steal even without the bot. A ledger file captures what fired, when, why, and how it ended; a learnings file has the agent state in plain language what each closed trade taught; optional personal notes before and after each position compound the trader's own judgment alongside the machine's. A live Bitcoin trade demonstrates the loop, including the warning that matters most: never let a single loss harden into a permanent rule, because even strong traders win a minority of the time and survive by sizing the wins.

Telegram setup closes the loop for life away from the desk: a bot created through BotFather, its token and chat ID handed to Claude, voice notes transcribed on the phone and forwarded as instructions. The catch is physical, and the video states it: the paired computer must stay on, which means a second Mac, a ten-dollar virtual server, or an old machine that never sleeps. The same plumbing is shown reaching stocks through Interactive Brokers, commodities and equities on supporting venues, plus portfolio briefings and property scans every morning, all framed as one mission of augmenting rather than replacing judgment.