A crypto arbitrage bot is a program that automates one or more parts of the arbitrage workflow: finding the spread, computing the route, alerting, and (sometimes) executing the trade itself. The fork that "arbitrage bot" ads almost never explain is what exactly the bot automates. A bot that finds and computes routes and sends you an alert is one thing. A bot that buys and sells on its own without you is another entirely. The difference matters because the arbitrageur's main enemy is non-working price gaps (phantoms), and blind auto-execution on them loses money faster than a human can blink. Below: how arbitrage bots work, what types exist, where their risks are, and why a "scanner + human" setup is usually safer than a blind automaton.

What an arbitrage bot is

Every arbitrage reduces to a route - the path of a single trade: asset + buy leg + sell leg + transfer. A bot automates part of that cycle. Which part defines its type and risk profile:

SCANNER ZONE · safe to automate risk lives here · 1 · Find scan books 2 · Compute net spread 3 · Alert ping operator 4 · Execute buy / sell / move human decides no final action money moves a scanner-alert stops at step 3 · an auto-executor crosses into step 4
Steps 1–3 only read and compute, so they are safe to automate. The whole risk profile flips at step 4, where the bot commits real capital.
  1. Find. Reads the order books of many exchanges and looks for a price gap. This is the most useful and safest part to automate - you can't do it by hand for the whole market.
  2. Compute. Subtracts fees and transfer cost, checks D/W statuses and depth - turns a raw spread into an executable net spread.
  3. Alert. Notifies the operator of a tradeable route (push, Telegram, a sound in the dashboard).
  4. Execute. Sends the buy/sell orders itself and kicks off the transfer.

The more steps are automated, the higher the speed - and the higher the cost of a mistake. A bot's safety and danger both live at step 4.

Types of arbitrage bots

By degree of automation, bots split into two broad classes.

Scanner / alert bot (find + compute + alert). Does steps 1–3 but doesn't trade on its own. Finds routes, computes the net spread, checks D/W and depth, filters phantoms, and sends the operator a finished list or alert. The human makes the execution decision. This is the class Finder belongs to.

Auto-executor (find + compute + execute). Does all four steps: sees a spread - buys, transfers, sells, all by itself. Attractive for its speed on liquid assets where a window lives seconds. But this is exactly where the main risks concentrate - because the bot executes phantoms too, unless they're filtered out perfectly.

Ads often sell an "arbitrage bot" as an auto-executor with a fixed yield. An honest heuristic: the higher the promised and "guaranteed" return of an auto-bot, the more likely it's either trading non-working gaps or isn't about arbitrage at all. Real arbitrage is a stream of small windows with real per-trade loss risk, not a fixed percent.

How a bot finds and computes a route

The useful part of any bot is finding and computing. Here the bot does exactly what a human does when reading a route by hand, only continuously and for the whole market: it computes the net spread (gross minus two takers), subtracts the network fee and time, checks that withdrawal and deposit are open on the specific network, and discards junk routes by depth and D/W. The layer-by-layer arithmetic of that computation is worked through in full in how to read a route, and why a status must never be guessed is in withdrawal windows and fees. The point here is one thing: the bot does all of it before any button is pressed.

gross spread → net edge, one route gross +4.0% − 2 takers −1.0% − transfer −2.1% net cutoff net +0.9% → tradeable net % after costs → D/W gate: withdrawal W🔴 closed → net forced to 0, route discarded
A 4.0% gross becomes a +0.9% net once two takers and the transfer are subtracted - and a single closed-withdrawal flag voids the route outright.

This part is pure upside: automating find and compute is safe, because the bot takes no final action. Risk appears only when you add automatic execution on top - which is what we look at next.

The risks of auto-execution

An auto-bot that trades on its own inherits all the risks of arbitrage - but without a human filter on the last step:

  • Bait gets executed blindly. A big percent usually isn't tradeable (why - in how to read a route). If the filter isn't perfect, the bot buys an asset it can then neither move out nor sell at the needed price. A gap like that would make a human suspicious. The automaton just hits "buy."
  • Stuck capital on a closed withdrawal. The bot bought on the cheap leg, but withdrawal there is 🔴 closed - money frozen in the asset. A human would see the flag and not enter. A bot without honest D/W enters (on this scenario - withdrawal windows).
  • Slippage at size. The bot sees the price at the top level, sends full size - but the book is thin and the real fill is worse. What lives at $200 doesn't live at $2000.
  • Convergence during transfer. While the asset moves over the network, the legs converge. The bot already bought at the old price and sells at the new one - the gap vanished in transit.
  • Technical failures. API latency, partial fills, leg desync - at an auto-bot's speed a mistake scales instantly.
  • Operations. Wrong network, memo, withdrawal limits - an auto-bot repeats the mistake on every trade until it's stopped.

Auto-execution does not make arbitrage "passive income." It's operational work with real per-trade loss risk, just at high speed. Without a perfect phantom filter and honest D/W, an auto-bot loses money faster than it makes it. Be especially wary of advertised "bots" promising a fixed return.

Why a scanner + human is often safer

Blind speed loses where the main job is to not trade junk. The "scanner finds and computes, human decides" setup is stronger for several reasons:

  1. The human is the last filter. The scanner cut 95% of the junk, but edge cases (odd D/W, a suspiciously thin book, an unfamiliar network) a human judges better than a blind automaton.
  2. Minute-long windows don't need milliseconds. On liquid BTC/ETH a window lives seconds and speed wins - but those routes are few and mostly taken by market-maker bots. The main flow for a retail arbitrageur is illiquid alts and cross-chain, where a window lives minutes. There a human with a good scanner keeps up.
  3. No risk of autonomous blowup. A scanner physically can't drain a deposit on a string of dud trades - it only shows. The final action is always the human's.
  4. Transparency. You see why a route is tradeable (net spread, D/W, depth), not "the bot said buy." That both teaches and protects.

That's why Finder is deliberately built as a scanner with alerts, not an auto-executor: it takes on the part impossible for a human (reading the order books of 20+ exchanges, computing net spread, pulling D/W, weeding out duds) and leaves the decision and execution to the operator.

Example: a minute of the scanner's alert log

The "scanner vs auto-bot" difference shows up best in the notification feed. Here's what an alert-bot's log looks like over one minute for token NEON - three events in a row, and in each the final action stays with the human:

12:04:01  NEON  KuCoin→Gate    net +1.6%  SOL  depth ~$5k  W🟢 D🟢   → ALERT to operator
12:04:18  NEON  MEXC→Gate      gross +7.4% but W🔴 (MEXC)  → suppressed, no alert
12:04:39  NEON  KuCoin→Gate    net +0.3%  (converged)       → retracted, window closed

What happened here:

  • 12:04:01 - a working window: net is positive after fees, both legs are open, the book holds ~$5k. The bot fires an alert and stops - whether to buy is the operator's call.
  • 12:04:18 - a big +7.4% gross on the same coin, but withdrawal from MEXC is 🔴 closed. The alert-bot suppresses it and doesn't ping the operator. An auto-executor without honest D/W would, on this same event, buy NEON on MEXC and freeze the capital - nothing to move out.
  • 12:04:39 - the first window converged to +0.3% within 38 seconds. The bot retracts the earlier alert. If the operator was just about to enter, they see the route is gone.

The scanner does the heavy part (found it, computed it, checked it, suppressed the junk, retracted the stale one) and hands the human a ready decision - instead of hitting "buy" blindly on the second event.

FAQ - crypto arbitrage bots

How does an arbitrage bot work?

It automates part of the cycle: finding the spread, computing the net route, alerting, and sometimes executing. The safe part is find and compute. The risky part is automatic execution, because the bot can blindly trade non-working gaps too.

Is an arbitrage bot legal and safe?

A scanner-alert by itself is safe: it only shows routes. The risk is in auto-execution - it inherits all the risks of arbitrage (false spreads, closed D/W, slippage) without a human filter. Be wary of ads for "bots" with guaranteed returns.

Do you need a bot for cross-exchange arbitrage?

For finding and computing - effectively yes: reading the order books of 20+ exchanges by hand is impossible, you need a scanner. For execution - not necessarily: on minute-long windows (illiquid alts, cross-chain) a human keeps up, and that's safer than a blind auto-bot.

Why is a scanner + human better than an auto-bot?

Because the main job is to not trade junk. The human is the last filter on edge-case routes, can't autonomously drain a deposit on a string of useless trades, and sees why a route is tradeable. An auto-bot's speed is only needed on a narrow class of liquid windows.

What exactly does Finder automate?

Find, net-spread computation, D/W checks, depth estimation and the weeding-out of non-working routes across 20+ CEX/DEX/perp venues, plus an alert. Finder deliberately leaves execution to the operator - it's a scanner with alerts, not an auto-executor.

Not financial advice. A bot is a tool, not a source of guaranteed income, and the fixed-percent "arbitrage" from ads usually isn't one. Auto-execution is especially treacherous: at speed an automaton replicates both profit and loss, and any window can close while the asset is in transit. Responsibility for running a bot is the operator's.

Where to next: the broad context - the pillar Crypto arbitrage guide. How to choose the scanner itself - spread screener. What a bot computes for you - how to read a route. Why a withdrawal is sometimes closed - withdrawal windows and fees. A scanner with alerts on live data - in the web dashboard.