Best AI agents for trading crypto, sorted by what the AI actually decides
"AI agent" on a crypto product page can mean a quant model that has run for years, a language model that writes a strategy from a sentence, or a chat window that suggests trades you place yourself.
The way to compare them is to ask three questions the marketing skips: what does the AI decide, who owns the strategy, and who holds the funds while it runs. The answers split the field into five kinds of product, each with its own risks, so "best" depends on which kind you need.
This page does that sorting, names the products in each group, and closes with how to evaluate any of them. It complements the perpetual-markets agent comparison, which compares execution models; this one compares what the AI is for.
Key takeaways
- Five kinds of "AI agent" exist: curated agents with a fixed mandate, intent-to-strategy builders, bring-your-own-model execution desks, agent marketplaces, and chat assistants.
- Ask what the AI decides. In some products it decides every trade; in others it writes a rule set once; in others it suggests and you click.
- Custody sorts the field faster than features. Exchange API keys mean exchange custody. A session key in your own wallet means the agent can trade and cannot withdraw.
- A backtest is not a track record. Any product showing simulated results should say so where you would act on them.
How the five kinds compare
| Kind | Examples | What the AI decides | Who owns the strategy | Where funds sit |
|---|---|---|---|---|
| Curated agent, fixed mandate | Deploy Finance (Income: Funding Rates, Superstar) | Every entry, hedge, exit within a defined strategy | The provider | Your own wallet; scoped session key |
| Intent-to-strategy builder | AKX AlphaOS, GT Protocol | Turns your plain-language intent into a strategy, backtests it, and may run it | You, via the AI | Self-custodial wallet (AKX) or exchange and Hyperliquid accounts (GT Protocol) |
| Bring-your-own-model desk | Hypertrade | Nothing; you bring the model. A risk engine approves, resizes, or vetoes its orders | You | Your account, with limited withdrawal authority |
| Agent marketplace | Avo (Solana), Neyro | Varies by agent; some are rule-based, some allocate across narratives | The agent's creator | Non-custodial per product; verify each |
| Chat assistant | Bybit TradeGPT, 3Commas AI Assistant | Suggests, explains, or configures; you execute | You | Wherever you trade |
Product descriptions are drawn from each company's public material at last review and can change. Verify status, custody, and permissions on the product's current documentation before funding anything.
1. Curated agents with a fixed mandate
A curated agent runs one defined strategy, end to end, and the provider is accountable for what that strategy is. You choose the agent and the allocation; the agent decides every trade inside its mandate. You do not write rules, tune parameters, or pick assets.
It suits someone who wants the result of a strategy they could not build and do not want to operate. In exchange, you accept the mandate as written and trust the provider's research as much as the venue.
Deploy Finance runs two live agents from a wallet you control. Income: Funding Rates is market-neutral: it holds a hedged position on Hyperliquid and collects perpetual funding, managing the hedge autonomously. Superstar is directional: it goes long, short, or flat on BTC, ETH, SOL, and HYPE perpetuals under a defined strategy. Both use a session key that can trade on whitelisted venues and cannot transfer or withdraw funds; you revoke it yourself. There is no subscription or fee on deposits; gas and trade fees apply.
The strategies are designed by a quant team rather than generated by a language model on request. That is a feature if you want a researched mandate, and a limitation if you want to describe your own.
2. Intent-to-strategy builders
These products put a language model between you and a strategy. You describe what you want in plain language; the AI compiles it into rules, backtests it, and in some cases runs it. You supply the intent; the AI decides how to express it.
Use one if you have a view and want it automated without writing code. The strategy is only as good as the sentence you gave it, and a model can produce a rule set that backtests well for reasons that will not repeat.
AKX presents AlphaOS, a natural-language interface that compiles intent into executable strategies from a self-custodial wallet, alongside a strategy marketplace. GT Protocol presents AI agents that build, backtest, and run strategies across Binance and Hyperliquid, with Telegram, Claude, and MCP control surfaces. For either, the custody question is which accounts the strategy runs on: a self-custodial wallet is one answer, an exchange API key is another.
3. Bring-your-own-model execution desks
Here the AI is yours. The product supplies the execution layer and a risk engine that can approve, resize, or veto what your model wants to do, with withdrawal authority held back from the model.
The fit is a technical trader who already has a model, an LLM-based agent, or a signal source and wants a controlled way to let it trade perpetuals. The product adds guardrails and takes no responsibility for the strategy, because the strategy is not theirs.
Hypertrade presents a trading desk for user-supplied AI or agents on Hyperliquid perpetuals, with a risk engine that gates orders and limits withdrawal authority. It is closer to infrastructure than to a product you allocate to.
4. Agent marketplaces
A marketplace lists agents built by others. Some are rule-based bots, some are LLM-driven allocators, some are portfolios under a theme. The AI decides whatever that agent's creator built it to decide. You pick the agent.
A marketplace suits someone who wants choice and will vet each agent themselves, because the marketplace does not. Quality varies widely, and the best and worst agents sit on the same listing page.
Avo presents a Solana-focused marketplace of AI trading agents and smart indexes, with an autonomous allocator, Miles, described as acting on natural-language requests and partly shown as coming soon. Neyro presents a non-custodial AI-agent execution layer with rule-based agents and user-controlled withdrawal. Both position on non-custodial control; verify the permission model of each specific agent before depositing.
5. Chat assistants
An assistant explains, suggests, and sometimes configures. It does not execute. You read its output and place the trade, or set the bot, yourself.
Pick an assistant if you want help thinking and intend to keep every decision. Here "AI-powered" describes the help you get; the trading stays with you.
Bybit's TradeGPT answers questions and surfaces ideas inside the exchange. 3Commas' AI Assistant configures DCA bots and runs backtests for you. Neither holds funds or places trades on its own; the 3Commas comparison covers the DIY model in detail. The How to use Claude for crypto page covers what a general assistant can do with the right tools connected.
What "AI" means in each kind
The word does different work in each group, and it is worth being precise before comparing returns.
- In curated agents, the AI is usually a quantitative model executing a researched strategy. Language models may sit in the interface, not the trade decision.
- In intent builders, a language model writes the strategy. The trading is then rule-based.
- In BYO desks, the AI is whatever you bring; the product's own intelligence is the risk engine.
- In marketplaces, it is whatever each creator claims.
- In assistants, it is a language model producing text.
If a product will not say which of these it uses, ask before you fund it.
How to evaluate any of them
Five checks, in order, before funding an AI trading product of any kind.
- Custody. Where are the funds while the agent runs, and can the agent withdraw or transfer them? An exchange API key with trade-only permission still leaves funds on the exchange. A session key in your own wallet does not.
- Decision boundary. Write down what the AI decides and what you decide. If you cannot, the product has not told you.
- Track record versus backtest. Simulated results should be labelled as such wherever you might act on them.
- Exit path. Revoke a permission, stop a bot and withdraw from an exchange, or redeem from a vault. Time it on a test amount.
- Fees and minimums. Subscription, performance share, spread, or nothing on deposits. Minimums tell you who the product is for.
Then the question that applies to every trading strategy regardless of the intelligence running it: what happens in the month the strategy is wrong, and who pays.
Verdict
If you want a researched strategy run for you from your own wallet, a curated agent; Deploy Finance is the live example, with a market-neutral agent and a directional one. If you have a view and want it automated, an intent-to-strategy builder. If you have a model, a BYO desk. If you want choice and will do the diligence, a marketplace. If you want help and intend to keep every decision, an assistant.
Learn more
- Best AI trading agents for perpetual markets
- How to use Claude for crypto
- Best automated trading platforms for non-traders
- How to choose an agent
- What are Deploy Finance agents?
- Risks
Start with Deploy Finance
Create a self-custodial Deploy Finance wallet and review the live agents.