On this page 11
Key takeaways
- Scores run from 1 to 5 and are a weighted average of five criteria. The weights are fixed and the same for every tool.
- Every tool gets one of five AI labels, from "Real ML" to "Rule-based", based on what the vendor documents, not what its homepage claims.
- Facts come from official pages and regulator registers, and every source carries the date we checked it.
- Vendor return figures and user counts are labelled as unverifiable claims, never stated as fact.
- No one can pay for a listing, a score or a ranking. Corrections are free and logged.
A score is only useful if you know what went into it. This page explains the rules behind every rating, label and comparison on AI Crypto Exchange, so you can decide how much weight to give them, and argue with us when you disagree.
Why we score at all
AI crypto tools are hard to compare. A grid bot, an on-chain analytics terminal and an autonomous agent solve different problems, and their marketing all uses the same words. A single number cannot capture that, but a consistent number with published rules is still more useful than a list ordered by whoever paid most.
So we keep scores narrow. They compare tools within a category, they are only given after a full review, and they are always shown next to the facts behind them. A tool can score well and still be wrong for you. Read the review before you trust the number.
The five criteria and their weights
Security and real AI carry the most weight because they decide whether a tool can hurt you and whether it is what it says it is. Pricing, track record and usability matter, but they rarely cost anyone their savings.
| Criterion | Weight | What we check | What pulls it down |
|---|---|---|---|
| Real AI | 25% | Is there learning, an agent, or only rules? Is the method documented? | Vague "advanced AI" claims, no technical detail |
| Security | 25% | API-key permissions, custody model, 2FA, incident history and response | Withdrawal-enabled keys, unexplained breaches |
| Pricing transparency | 20% | Published plans, real free tiers, total cost with fees | Hidden volume fees, sales-only pricing for retail |
| Track record | 15% | Years operating, public team, regulatory status, complaint patterns | Anonymous team, regulator warnings |
| Usability | 15% | Setup time, demo or paper trading, apps, documentation | No demo mode, thin or outdated docs |
Real AI (25%)
We read the vendor's documentation, not its landing page. A tool scores highly when it tells you which models it uses, what data they learn from and how often they retrain, as Freqtrade does for FreqAI, which is described in a peer-reviewed paper. An honest rule-based tool sits in the middle. The lowest marks go to products that claim AI and cannot explain it. Regulators take this seriously too: in March 2024 the SEC settled charges against two investment advisers over false AI claims, with $400,000 in combined penalties.
Security (25%)
For bots, the first question is what an API key can do. A tool that works with trade-only keys and blocks withdrawal permissions, as Bitsgap does, starts ahead. We also look at whether funds are custodial, whether 2FA is available, and how the company handled past incidents. The 2022 3Commas API-key leak still weighs on its score because of how it unfolded. For exchanges we add cold storage, published certifications and major hacks, such as the February 2025 Bybit breach of roughly $1.5 billion.
Pricing transparency (20%)
We price the whole thing: subscription, trading fees, profit shares and any surcharge on larger balances. Stoic, for example, adds 5% a year on balances above $10,000, which matters more than its headline plan price. Clear public pricing with a usable free tier scores well. "Contact sales" for a retail product does not.
Track record (15%)
How long has the company operated, is the team public, is it registered where it serves customers, and what do user complaints say? We check review platforms such as Trustpilot and note the date, because ratings move. A regulator warning or a sudden platform shutdown, like the end of 3Commas v1 in September 2026, lowers this score.
Usability (15%)
Can a reasonably careful person set it up in an afternoon, test it without real money, and find answers in the documentation? Mobile apps and paper trading help. A tool built for developers is not marked down for needing code, but we say clearly who it is for.
How a score is calculated
Each criterion gets a sub-score from 1 to 5 in half-point steps. We multiply each by its weight, add them up and round to one decimal. Here is a hypothetical example, not a real tool:
Hypothetical worked example of the weighting
| Criterion | Weight | Sub-score | Weighted |
|---|---|---|---|
| Real AI | 25% | 0.75 | |
| Security | 25% | 1.13 | |
| Pricing transparency | 20% | 0.80 | |
| Track record | 15% | 0.53 | |
| Usability | 15% | 0.68 | |
| Overall | 100% | 3.875, shown as 3.9 |
Roughly, 4.5 and above means outstanding for its category, 4.0 to 4.4 is a strong choice, 3.5 to 3.9 is good with real caveats, and 3.0 to 3.4 means use with care. Below 3.0 we would not suggest the tool to most readers.
Two rules override the arithmetic. A tool that promises guaranteed or fixed returns is not scored at all; it belongs in our AI trading scams guide. And an unresolved security incident caps the overall score until the company explains what happened and what it changed.
The five-level AI reality scale
Alongside the score, every tool gets one label that describes what its AI actually does. It is the most useful single filter we offer, because "AI" on a sales page can mean anything from reinforcement learning to a chatbot bolted onto an old grid bot.
| Label | What it means | Examples | The question we ask |
|---|---|---|---|
| Real ML | Trains or retrains machine-learning models on market data | Freqtrade with FreqAI | Which models, trained on what data, retrained how often? |
| AI agent | An LLM agent that can plan and execute actions | Nansen AI, Bankr, OKX Agent Trade Kit | What can it do without your approval, and what limits can you set? |
| LLM assistant | A chat assistant for research, explanations or strategy drafting | LunarCrush, TrendSpider Sidekick, Bybit TradeGPT | Where does its data come from, and can it place trades? |
| Rules + AI layer | Rule-based bots with an AI helper on top (assistant, optimiser or MCP) | Bitsgap, Pionex, Coinrule | What does the AI part change, and is the method published? |
| Rule-based | Automation without machine learning, often marketed as AI | Gunbot, TradeSanta, Bubblemaps | Is it marketed as AI anyway? |
We assign the label from the vendor's own technical documentation, help pages and, for open-source tools, the code. When the documentation is silent, we choose the lower label and say so in the tool's note. A vendor can move up a level by publishing how its system works. Our guide to how AI crypto trading bots work goes deeper into what sits behind each label.
How we verify facts
Every price, date, licence and incident on the site traces back to a source we can name. The order we work in:
Start with the official page
Pricing, features, supported exchanges and country restrictions come from the vendor's own pricing page, documentation or terms. We record the URL and the date we read it.
Check regulator registers
Licence claims are checked on the regulator's own register: NMLS Consumer Access for US money transmitters, the FCA register in the UK, and ESMA's MiCA lists in the EU. Regulator warning lists are checked too.
Look for incidents and complaints
We search for hacks, leaks, enforcement actions and shutdowns, preferring the company's own statement and reputable reporting. Review-site ratings are noted with the date we saw them.
Test what we can
Where a free plan, demo mode or paper trading exists, we use it to confirm setup steps, API-key options and what the AI feature actually produces.
Date it and label the gaps
Each fact is published with a check date. Anything we could not confirm is either left out or clearly labelled as a vendor claim.
We use AI models to speed up the reading: summarising long documentation, comparing plan tables or flagging changes since our last check. A person then checks each extracted fact against the original page before it is published. If the model and the source disagree, the source wins.
Claims we cannot verify
Some numbers simply cannot be checked from outside a company. Bot returns, win rates, "users served" and assets under management are the common ones. We treat them like this:
- Returns and win rates are not used in scores and are not repeated as fact. If we mention one, it is labelled "vendor claim, not independently verifiable".
- User counts and country counts are attributed to the company ("the company claims 15M+ users"), never stated in our own voice.
- Backtests are described as backtests. A strategy that worked on past data can fail live, and overfitting is the most common reason.
Corrections policy
When we get something wrong, we fix it. Anyone can report an error, reader or vendor, by emailing contacts@ai-crypto.exchange with the page and a source. We re-check the fact against the primary source. If we were wrong, we correct the page, update the check date and, where the change is material (a price, a licence, a security fact or a score), add a short note on the page saying what changed and when.
If a correction affects a criterion, we re-calculate the score using the same weights. We do not quietly delete critical findings because a company asked; we only change them when the evidence changes.
Update cadence
All facts on the site were checked in October 2026. After that, three things trigger a re-check: a vendor announcing a pricing or product change, a regulator or security event, and a reader report. On top of those, we work through every review and the full directory on a regular cycle, with fast-moving categories such as AI agents and exchanges checked more often than open-source frameworks. The "Updated" date at the top of each page shows when it was last reviewed.
Conflicts of interest
No tool vendor, exchange or token project can pay to appear on this site, to improve a score, or to be ranked higher. We do not accept free premium plans in exchange for coverage. Team members do not review products they have a personal financial stake in beyond ordinary use, and they do not write about tokens they are actively trading.
Some pages suggest an exchange for buying and holding crypto. We choose venues whose licences can be checked on public registers, and that suggestion is kept separate from scoring: it never changes a tool's score or position in a ranking. If you think we have fallen short of any of this, tell us. Our about page explains who we are, and the disclaimer covers what our content is and is not.
Frequently asked questions
Who decides the scores?
Scores are set by our editorial team using the five weighted criteria on this page. Every sub-score is tied to dated, linked evidence, so anyone can re-check how we got there. AI models help us gather and summarise documentation, but they never assign a score.
Can a company pay for a better score?
No. Vendors cannot pay to be listed, reviewed, scored, ranked or removed, and we do not publish sponsored reviews. A company can send us evidence that a fact is wrong, exactly as a reader can. If the evidence holds up, we correct the fact and re-score if the change affects a criterion.
Why do some tools in the directory have no score?
We only score tools we have reviewed in full, currently 12 of 55. Other tools are listed with checked facts and an AI label, but without a number. Showing a score we had not earned would undermine the ones we have.
Do you test bots with real money?
Where a product offers a free plan, demo or paper-trading mode, we use it to check setup, key permissions and how the interface behaves. We do not publish profit figures from our own tests as evidence that a bot works, because a few weeks of results in one market tell you very little about the future.
How do you score a tool with no AI at all?
It can still score well on security, pricing, track record and usability. On the Real AI criterion, a tool that is honestly described as rule-based scores in the middle of the range. The lowest marks go to tools that claim AI they cannot explain, because that is a transparency problem as well as a technical one.
Keep reading
Sources and fact-check dates
We link to primary sources wherever possible. Dates show when the page was published or when we last checked it.
- CFTC customer advisory: AI Won't Turn Trading Bots into Money Machines · checked 2026-10-21
- CFTC Release No. 8854-24 (25 January 2024) · checked 2026-10-21
- SEC Press Release 2024-36: charges over false AI claims · checked 2026-10-21
- FCA warning: Quantum AI · checked 2026-10-21
- NMLS Consumer Access · checked 2026-10-21
- FCA Financial Services Register · checked 2026-10-21
- FreqAI paper, Journal of Open Source Software (2022) · checked 2026-10-21
