Finding Arbitrage Opportunities

Overview

Compare one exact pair's best bid and ask across nine exchanges and compute every cross-venue spread, gross and net of stated taker fees. A triangular search checks three-trade cycles inside one venue.

Venues: Kraken, Coinbase, KuCoin, Gate, OKX, Bitfinex, Crypto.com, Binance, Bybit. Binance and Bybit refuse US networks (HTTP 451/403); the tool names them and carries on.

Pairs with different quote assets are never compared: BTC/USD against BTC/USDT measures the USDT/USD rate, not an arbitrage.

Prerequisites

• Python 3.9 or newer. Standard library only: no packages, no API keys.
• Network access to the exchanges' public endpoints.

Instructions

All commands run python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py. Every command accepts --format json, --fee VENUE=PERCENT (the actual account taker fee), --demo (bundled recorded data, marked on every line) and --timeout SECONDS.

1. Scan one pair across venues:

   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py scan BTC USDT
   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py scan ETH USD --venues kraken,coinbase,bitfinex,gate

1. Set actual account fees (fees vary by tier, region and token discounts):

   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py scan BTC USDT --fee coinbase=0.6 --fee kucoin=0.1

1. Express the best net spread in money for a trade size in base units:

   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py scan SOL USDT --amount 25  # 25 = trade size in SOL

1. Search triangles on one venue (okx, kucoin, gate), using pairs above a 24-hour volume floor:

   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py triangular okx
   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py triangular kucoin --fee kucoin=0.1 --min-volume-usd 250000  # 250000 = USD volume floor

1. Watch a pair for a few minutes (minimum interval 5 seconds):

   python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py monitor BTC USDT --interval 15 --count 8 --threshold 0

1. See fee assumptions and their sources: arb_finder.py venues.

When reporting, give the spreads with the fee assumptions and the "not included" list below. Describe what the data shows; never recommend a trade.

Output

Console tables end with sources and the line "Research output only. Not financial advice. This tool places no trades." JSON output is one envelope with skill, version, command, generated_at, mode (live or demo), sources, failures, data, not_included and disclaimer.

Every spread excludes withdrawal and network fees, the time to move funds between venues, price moves in that time, and depth beyond the best bid and ask. A positive net spread at the top of the book is not a collectable amount.

Fee assumptions are base-tier spot taker fees taken from each venue's published schedule (dates and sources in arb_finder.py venues). Gate's is read live from its API. KuCoin and Coinbase do not show a base-tier fee without a login, so their net column reads n/a until --fee sets it. Details: references/implementation.md.

Error Handling

More in references/errors.md.

Examples

Cross-venue BTC/USDT check (an excerpt of captured live output; full runs with their capture time are in references/examples.md, and live figures change every minute):

python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py scan BTC USDT --top 3

Buy        Sell            Gross  Net of fees   Size at best
bitfinex   okx          +0.0293%     -0.3208%    1.35507 BTC
bitfinex   coinbase     +0.0291%          n/a     0.0109 BTC
bitfinex   cryptocom    +0.0258%     -0.4743%     0.3817 BTC
No venue pair stays positive after assumed taker fees.

Triangles on OKX as JSON:

python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py triangular okx --format json

Try it offline:

python3 ${CLAUDE_SKILL_DIR}/scripts/arb_finder.py scan BTC USDT --demo

Resources

• references/implementation.md: endpoints, formulas, fee sources, what is not covered
• references/examples.md: real output for every command
• references/errors.md: error messages and fixes
• Tests: python3 -m unittest discover -s tests -v (offline)
• Refresh the demo recording: python3 scripts/record_demo.py