Crypto Trading Bots

What this category is

Software that places orders without you watching, and the libraries that software is built on. Two kinds of buyer share the page: somebody who wants a strategy running unattended, and somebody who is writing that strategy's plumbing and needs one client that speaks to forty venues.

The boundary against backtesting libraries runs through what the tool is for, not what it can do. Almost everything here can backtest, and almost everything there can place an order. A framework people install to find out whether an edge survives belongs there; one they install to keep a market-making quote alive belongs here.

The split that decides everything else

Every card on this page falls on one side of a single line, and it is a custody line rather than a technical one: does somebody else hold a key that can trade your account?

Hosted, keys on the vendor's servers. 3Commas, Cryptohopper, Bitsgap, OctoBot Cloud, WunderTrading, Coinrule, CoinAPI EMS and the cloud edition of HaasOnline. Two of those are now one engine under two brands: 3Commas switched v1 off on 11 September 2026, and v2 was built in partnership with WunderTrading. You paste an exchange key into a web app, the vendor's infrastructure runs the strategy whether or not your machine is on, and the key sits in their database for as long as the subscription lasts. That is not a flaw — it is the entire product, and it is the only way a bot keeps quoting while your laptop is shut.

Self-hosted, keys on your machine. CCXT, ccapi, Freqtrade, Hummingbot, OctoBot, Passivbot, GoCryptoTrader, Gunbot, dYdX v4 Clients, the Hyperliquid Python SDK, Superalgos and HaasOnline's self-hosted TradeServer. Nobody else can reach the account, and nobody else is keeping it up either. The self-hosted list is the same cut taken across the whole catalogue. Check the dates before choosing one: Superalgos's last stable release is from November 2024, and on 2026-10-04 its domain showed an expired-registration notice.

Self-hosted is not the same claim as open source, and three cards here prove it. Gunbot is a signed binary you install and register against a wallet address — you own the deployment and cannot read or fork what is executing your strategy. Since October 2026 new buyers get it only as Gunbot Unlimited, 29 euros a month; the lifetime licence is sold only as an upgrade to existing holders. HaasOnline's TradeServer is the same case at a far higher monthly price: it runs on your hardware and its source is closed. dYdX v4 Clients publish their source and ship it under a custom licence, not the AGPL their README badge and package metadata claim, whose grant terminates automatically if you fall out of step with the venue's terms of use. Running it yourself answers the custody question and leaves the audit question open.

What the money is when the product is free

Ten of the nineteen cards on this page cost nothing at all, and almost none of them is unfunded. Five mechanisms do nearly all of the paying, and who pays for your free trading bot works through each one and how to find it in a vendor's own documentation in about ten minutes. Two of them are worth naming here because they are visible on the cards.

A builder fee is the one that comes out of your fill. CCXT adds one basis point on top of the exchange's own fees on seven DEX venues, on by default, switched off with a single option. It is documented and it is still a charge most readers will not have priced, because it does not appear on any pricing page — there is no pricing page. The Hyperliquid Python SDK, by contrast, sets no builder unless you pass one; it ships the approval call for people charging their own users. WunderTrading is the hosted case, and it prints the fee on its pricing page: on Hyperliquid its Premium plan carries no subscription and charges a 0.035% builder fee on each trade instead, a cost that scales with volume where a subscription does not.

Everything else comes out of fees you were already paying. Exchange API-broker rebates fund CCXT; the Hummingbot Foundation is paid by venues for maintained connectors and sponsorships, and four of the referral links in its README advertise a fee discount to you as well; OctoBot names nineteen exchange partners; 3Commas routes every sign-up button on its exchanges page through affiliate codes. None of that is hidden and none of it is disqualifying. It is simply the reason a particular venue is the default in the setup guide, and it is worth knowing before you conclude that the default was chosen on latency.

Two cards appear to take nothing at all. Barter and GoCryptoTrader carry no referral links, no sponsor banners and no broker programme in their READMEs — GoCryptoTrader asks for a Bitcoin donation instead. On a page where the normal funding model is the venue, that is unusual enough to be worth stating.

The headline prices do not compare

The hosted platforms cluster within a few dollars of each other at the entry tier and are metering three different things: bot types on one, a single bot whose effort scales on another, bot counts per type on a third. Two people with identical strategies land on different vendors, and the plan you need is decided by how many positions you intend to have open at once and in how many shapes — not by which strategies you want. 3Commas vs Cryptohopper vs Bitsgap converts all three into the same unit; OctoBot vs Passivbot does the equivalent job on the self-hosted side, where the price is zero and the thing being metered is your attention.

One card is not in that market at all. CoinAPI EMS is an institutional execution API billed by the account-hour rather than by the month, with TWAP and VWAP over twenty-six destinations — the shape of product the trading terminals page is full of, filed here because it is bought to be driven by code.

The connector layer is the load-bearing part

Most of this category does not talk to exchanges itself. It depends on a connector library that normalises forty REST dialects into one call, and that dependency is invisible on the feature page. It matters for two reasons.

The exchange count you are reading is usually the connector's, not the product's. A framework advertising a hundred venues is counting what its dependency supports, and support in that list means "an adapter exists", not "somebody traded through it last month". Freqtrade is the honest version of that sentence: it reaches exchanges through ccxt, its documentation names twelve spot entries and seven futures venues as supported, and of everything else ccxt can reach it says it cannot guarantee they will work.

The connector's maintenance status, not the framework's, decides whether your venue keeps working. A venue changes an endpoint, the connector lags, and the framework you chose for its active development is suddenly unable to reach the exchange holding your balance. It is also where the rate limit is enforced or not — a connector that does not throttle on your behalf will get the key banned by the venue rather than by the framework.

What else actually decides the choice

What the paper-trading mode actually simulates. Dry-run against live prices and backtest against history are different claims, and neither models what your own order does to the book. Freqtrade's dry-run mode runs the same strategy class its backtester runs and its live bot trades, so what you watch is the code you will deploy — but its automatic pairlists do not run in a backtest, so the static list you tested is not the universe the live bot picks.

Whether it models perpetual funding at all. Funding is paid every few hours and is one of the two largest cash flows on a leveraged position. An engine that never charges it will show an automated perp strategy earning money it would not have earned, and the same defaults problem set out on backtesting libraries applies to the backtester bolted into a bot — usually more so, because it was not the reason the product was built. Freqtrade is the exception worth knowing: in futures mode it downloads mark and funding-rate series and settles funding during the backtest, where the venue serves funding history; where it does not, the fallback is a constant the docs recommend setting to zero. Freqtrade vs Jesse sets that against a framework that charges no funding in a backtest at all.

Whether your own code can reach it. A hosted platform with no API, no webhooks and no TradingView execution is a closed box however good the bots are, and that is a real position several vendors take. If your stack has to talk to the platform, it eliminates candidates before anything else is considered.

What is not here

Products that sell signals, calls or managed allocation. This catalogue lists software people run against their own keys and their own decisions; a subscription that tells you what to buy is advice, and the duty that comes with it is not one an unaffiliated catalogue can carry.

Copy-trading features inside a bot are a capability flag, not an exclusion — the line runs through whether the vendor's business is publishing the calls, not through whether the software can follow somebody else's.