Skip to content

Repository files navigation

iBitLabs — Public Mirror

Curated public mirror of the iBitLabs live trading experiment. Canonical live surface: https://www.ibitlabs.com

What this is: a non-coder (Bonnybb, in crypto since 2017, zero programming background) used AI to build a mean-reversion crypto trading bot on Coinbase SOL perpetual futures in about seven days, funded it with $1,000 of her own real money, and runs it in public. Every trade auto-posts to a Telegram channel. The full live experiment — balance, open positions, PnL, fees, funding, trade log, indicator values — is exposed as a JSON feed at https://www.ibitlabs.com/api/live-status.

This repository is the public mirror of the experiment's infrastructure. It contains the executor, the reconciler, the shadow-rule instrumentation, the daily chronicle CMS, and the analysis scripts — without the live API credentials or the live trading account's instance state. Fork it, read it, learn from it. Running it against your own Coinbase account is possible; running it sensibly requires reading the docs first.

Status snapshot (as of last commit, not live): Day 36 of live trading. Account ~$984 (-1.6%). 61 trades. 54.1% win rate. SOL trades live; ETH paper-mode in parallel through 2026-05-20. For the live numbers, see https://ibitlabs.com/signals or hit the /api/live-status JSON feed.

📖 Read the long-form essay: The "We" of AI Co-Founding — 9,500 words, 22 minutes. The full thinking behind this repo: why $1,000 is the smallest unit where attribution carries weight, what "AI as co-founder" means operationally, and how Mode 1 / Mode 2 / Mode 3 differ. Five concrete observations from 26 days of running this in public, ten chapters across three acts.

Why this repo exists

Most "AI-built trading bot" demos are on paper accounts with cherry-picked screenshots. iBitLabs is different, and the public mirror carries that intent forward:

  • Real money, real exchange. The executor in this repo is the same code that placed the last trade on the live @ibitlabs_agent Coinbase account, versioned and readable.
  • Open failure mode documentation. The repo's docs/ directory documents specific real production bugs (ghost position bug, fee cushion miscalibration, narrow-window backtest retraction) in the prose-equivalent of post-mortems.
  • Instrument-before-rule observability pattern. See docs/shadow_12h_rule.md and scripts/analyze_shadow_12h_rule.py. A small reusable pattern for shipping a rule's write-side (logging what-would-have-happened) without its act-side, collecting 30 days of evidence, then deciding from data.
  • Transparent retractions. When a claim turns out wrong, the retraction is a first-class artifact. The canonical example is the public walk-back of a 90%-win-rate claim after the 13-month backtest collapsed to −46%. That retraction has since been folded into the serial novel at https://ibitlabs.com/saga/en.

What's in this mirror

.
├── sol_sniper_executor.py      # position open / manage / close, incl. shadow rule
├── sol_sniper_config.py        # strategy config (see disclaimer below about numbers)
├── state_db.py                 # SQLite trade log + MFE/MAE + shadow persistence
├── scripts/
│   └── analyze_shadow_12h_rule.py  # post-hoc EV analysis of shadow log
├── receipt-rule-engine/        # reactive + self-healing layer (v0.2, added 2026-05-13)
│   ├── engine.py               #   tails receipt JSONL chains, fires rules
│   ├── rules.example.py        #   copy → rules.py, edit chains + actions
│   └── scripts/                #   reference Tier-1 actions (anchor, reconcile)
├── receipt-viewer/             # local HTML dashboards for chain observation
│   ├── index.html              #   chain explorer
│   ├── dashboard.html          #   3-stream KPI dashboard
│   └── monitor.html            #   rule-engine output monitor
├── docs/
│   ├── shadow_12h_rule.md      # 30-day observation window spec
│   ├── days_cms.md             # bilingual daily-chronicle CMS operator guide
│   └── live-restart-checklist.md  # production bot restart / recovery procedure
└── SKILL.md                    # cryptoskill.org registry manifest

New as of 2026-05-13: receipt protocol layer

The bot now emits a hash-chained receipt for every action — entry, exit, reconcile, heartbeat — to audit_export/*.realtime.receipt.jsonl. Two layers consume that chain locally:

  • receipt-rule-engine/ — a small daemon that tails one or more chains and fires actions when events match declarative rules. Supports the absent pattern (fire if no event of kind X in Y seconds), auto: true/false gating on shell actions (default off; ntfy/iMessage low-blast-radius and never gated), and engine self-heartbeat. Rules are data, not code paths. Tier discipline documented in engine.py.
  • receipt-viewer/ — three static HTML pages served from 127.0.0.1 (no server framework, just file:// or any HTTP). index.html is a chain explorer; dashboard.html is a 3-stream live/shadow/paper KPI surface; monitor.html watches the rule-engine output chain. Open the HTML files directly or python3 -m http.server 8092 -d receipt-viewer/.

The protocol itself is its own pip-installable package — see bbismm/receipt. This mirror bundles the engine + viewer because they're how we use the chain in practice; you can swap in your own consumers.

What's NOT in this mirror (and why)

  • Live Coinbase API credentials (obvious). Set your own in .env — never commit.
  • Live account state file (sol_sniper_state.json). Instance-specific.
  • Trade DB with actual live fills (sol_sniper.db). Instance-specific.
  • Launchd plists + bot scripts with env-var hardcoded secrets. Deploy tooling is yours to write.
  • Moltbook + Telegram + Notion automation code. That's brand / content infrastructure, not trading infrastructure. Lives in the private main repo.
  • The website code (web/). Deployed on Cloudflare Pages, separate concern.

About the strategy numbers

These numbers are mine, not yours. sol_sniper_config.py contains the exact tuning values in use on the live iBitLabs account: 5% stop-loss, 1.5% trailing-stop activate, 0.5% trailing drawdown, 80% position sizing, StochRSI thresholds around 0.10/0.90, regime-window 288h, and so on. These are documented for reproducibility of the experiment itself (so anyone reading a day's essay can see exactly what the bot was doing that day), not as a template for forking.

Do not fork these numbers verbatim and expect them to work on your account. They were arrived at through ~20 iterations of backtests and live calibration over the first 17 days of the experiment. They reflect one particular strategy version (hybrid_v5.1) trading one particular instrument (Coinbase SOL PERP SLP-20DEC30-CDE) during one particular market regime (SOL down ~3% 30-day trailing as of commit). Run your own sweep. Read the retraction essay before you trust any backtest window narrower than twelve months.

Install + first run

See STARTER.md — clone, install deps, run python paper_quickstart.py, see a full trade lifecycle in under a minute. No Coinbase API key needed for the quickstart (it uses a mock exchange).

For wiring the executor to live data later, see the "Wiring to live (later, with care)" section of STARTER.md.

Requirements (minimal set, in requirements.txt): coinbase-advanced-py, pandas, numpy. No ML frameworks, no heavyweight TA libs.

Integrations for agents without running the bot

If you only want data, no integration:

  • GET https://www.ibitlabs.com/api/live-status — live JSON feed, no auth. Use cases: RAG context, cited examples in agent responses, comparison reference.
  • https://ibitlabs.com/saga/en — bilingual serial novel narrating the experiment day by day, dual-POV prose form. Vol 2 publishes one new chapter every night at 22:30 EDT against real trade data. Good source material for "what does a non-coder's AI-built trading bot look like day by day."

An MCP server wrapper is in the mcp-server/ directory (when shipped).

Related links

License

  • Code: MIT
  • Prose / documentation: CC BY 4.0

See LICENSE.

Maintainer

Bonnybb · GitHub @bbismm

Issues welcome. The experiment is explicit about not selling anything — there is no paid tier, no signal service, no course. The most meaningful way to engage is to read the docs, fork the pattern (not the numbers), and tell me what breaks.

About

Curated public mirror of the iBitLabs live trading experiment — executor, reconciler, shadow-rule scaffold, no creds.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages