iBitLabs — Live AI-Built Crypto Trading Bot

A non-coder's public live experiment. Mean-reversion trading bot on Coinbase SOL perpetual futures, built with AI assistance (primarily Claude), running with $1,000 of real money since 2026-04-07. Every trade auto-posts to a public Telegram channel. The full source code, strategy versions, shadow-rule instrumentation, and daily dual-POV chronicle are public.

• Category: trading, exchanges (Coinbase perp), ai-crypto
• Author: Bonnybb (single creator; in crypto since 2017; zero prior programming background)
• Status: Community (official for the ibitlabs_agent brand account). Live since 2026-04-07. Day 17 at time of SKILL.md.
• License: Code under MIT. Content under CC BY 4.0 (see LICENSE).
• Repo: https://github.com/bbismm/ibitlabs
• Live surface: https://www.ibitlabs.com
• Live JSON feed: https://www.ibitlabs.com/api/live-status

What it does

An end-to-end live crypto trading system built in ~7 days by a non-coder using AI:

1. Trading executor (sol_sniper_executor.py): opens / manages / closes positions on Coinbase SOL perpetual futures. Uses a mean-reversion entry (StochRSI oversold + Bollinger Band mid-touch + regime filter), tiered take-profit, stop-loss, and a trailing stop.
2. Exchange-truth reconciler (com.ibitlabs.db-exchange-reconcile): every 15 minutes, diffs local SQLite state against the Coinbase API. Flags drift. Caught the "ghost position bug" (reduce_only flag missing on close orders) in production.
3. Shadow-rule instrumentation (see docs/shadow_12h_rule.md): evaluates a hypothetical compound exit rule every tick and writes a JSONL log line when fired, without executing. Used to collect 30 days of observational evidence before shipping any rule change to live execution.
4. Live-status JSON API at /api/live-status exposing balance, open position, PnL, fees, funding, trade count, win rate, indicator values, and strategy version — updated on every bot tick.
5. Auto-post Telegram channel (@ibitlabs_sniper): every entry, exit, PnL, and fee is posted within seconds of exchange execution.
6. Public dashboard (/signals), daily chronicle (/days, bilingual EN+中文, dual-POV), academy (/academy), and essays (/essays).

Why this is a skill worth installing / studying

Most "AI-built trading bot" projects are demos on paper accounts with cherry-picked screenshots. iBitLabs is different:

• Real $1,000 on Coinbase. Every trade is verifiable against the exchange.
• Open-source executor code — not a black-box "proprietary AI" wrapper. You can read the file that placed the last trade.
• Public failure mode documentation. The repo's docs/shadow_12h_rule.md, docs/days_cms.md, and essays all document specific bugs (ghost position bug, fee cushion miscalibration, narrow-window backtest trap) in the prose-equivalent of post-mortems. Useful as teaching material for other AI-agent developers building execution systems.
• Instrument-before-rule pattern (Python). The shadow-rule pattern is a generalizable observability technique for any decision system: ship the write-side (logging) before the act-side (execution), collect evidence, decide from data. See scripts/analyze_shadow_12h_rule.py.
• Transparent retractions. The repo's git log contains the commit that retracted a public 90%-win-rate claim after a 13-month backtest collapsed to -46%. Public wrongness is version-controlled.

Install / run

git clone https://github.com/bbismm/ibitlabs.git
cd ibitlabs
pip install -r requirements.txt
# Copy .env.example to .env and fill in Coinbase API credentials
cp .env.example .env
# Paper mode for safety first:
python3 sol_sniper_main.py --paper
# Live mode (uses your Coinbase futures_buying_power):
python3 sol_sniper_main.py --live

The bot runs as a long-lived process. For production deployment on macOS, see the launchd plists in launchd/ (the ones used on the live account). On Linux, systemd unit files need to be written separately.

Integration surfaces for other AI agents

You do NOT need to run the bot to get value. Two lightweight integration paths:

1. Read live data — GET https://www.ibitlabs.com/api/live-status returns a JSON document with the full live state. No auth required. Rate-limited at the edge. Use cases: citing real live trading data in AI-agent responses, building agent dashboards, building "compare against a live reference" tooling.
2. Read the daily chronicle — GET https://www.ibitlabs.com/days (HTML) or fetch data/days.json directly. Bilingual dual-POV daily prose chronicle of a live AI-built trading bot. Use cases: RAG context for agent training, cited examples of AI-assisted non-coder development, narrative examples of failure + recovery.

A dedicated MCP server wrapper is planned (see mcp-server/ in this repo).

Environment variables

All via .env (gitignored). Never committed:

• CB_API_KEY — Coinbase Advanced Trade API key (futures-enabled)
• CB_API_SECRET — EC private key PEM for Coinbase
• TELEGRAM_BOT_TOKEN — optional, for auto-post feature
• TELEGRAM_CHAT_ID — the channel ID to post to

web/wrangler.toml references Cloudflare Pages secrets (never plaintext keys).

Safety and status of this skill

Do NOT treat this as investment advice or signal source. The experiment's explicit thesis is the opposite — that a non-coder + AI can build a trading system at all, not that it will make money. As of Day 17 the account is down ~2.6% net, with carry cost (fees + funding, $57.64) exceeding realized trading losses ($14.02).

The code is a working reference, not a production-ready framework. If you fork, read docs/ first and understand the reconciler logic before you trust local state for anything.

Maintained by

Bonnybb · Lives on GitHub as bbismm. Contact via GitHub Issues on this repo, or Telegram channel @ibitlabs_sniper.