Agentic trading

Where the line sits

The term is used loosely enough that it is worth saying what is being counted and what is not. These are the boundaries this page applies, applied the same way to every product in the table.

• Counted: the agent can cause a real order

  On a real venue, with real money. The order may still need a confirmation, and the product is counted either way as long as the agent is what submits it.

• Counted, and labelled: the agent drafts and stops

  A person submits. Kept in the table rather than excluded, because several of these are described as agentic trading elsewhere and a reader deserves to see where they differ.

• Not counted: reading prices

  A server that returns quotes, charts, balances or research is a market data server. Most of what is listed in MCP directories under trading is this, and counting it would make the field look several times larger than it is.

• Not counted: a bot you configured

  A rules engine, a copy-trading product or a scheduled strategy does the same thing every time. No model is choosing at the moment of the trade, which is the thing that makes this category its own question.

• Not counted: paper trading

  A simulated account is a useful place to test an agent and it is not this. Nothing here counts unless real funds are reachable.

Three questions that separate these products

Almost every practical difference between one of these products and another comes down to three questions. None of the three has a better answer than the others, and this page does not suggest one. They are the axes because they are what changes for you, not because one end of each is preferable.

1. Who supplies the agent

The axis a reader usually cares about first, because it decides whether the assistant they already use is the one that trades. 22 of the 24 let you bring your own, and the remaining 2 do that as well as shipping an agent of their own. Nothing here supplies only its own agent. That is the finding of this axis, and it stayed hidden until a review found one row filed against a description that was not in its cell.

• 22

  You bring the assistant

  You already use an AI assistant. The venue publishes a server or connector, you point your assistant at it, and the assistant you brought is the one that trades.

  Alpaca MCP Server, Binance Agent OS, Bitpanda Fusion MCP, Bybit MCP Server, Coinbase for Agents, Composer MCP, cTrader AI Agent Connect, eToro Agent Portfolios, Fensory, Gate MCP, Interactive Brokers AI Integration, Kraken CLI, MetaMask Agent Wallet, OKX Agent Trade Kit, Open-source Hyperliquid servers, Phantom MCP, Robinhood Agentic Trading, Scalable Capital Agentic Investing, tastytrade MCP, TradeStation MCP, Trading 212, Walnut

• 0

  The venue supplies the agent

  The venue supplies the agent. You configure it inside the venue’s own product and there is no documented way to drive it from an assistant of your own.

  None

• 2

  Both

  The product does each, usually as separate offerings that share a brand. Which one you have changes what it can do, so the name alone does not tell you.

  Senpi, Webull

• 0

  Not stated by the vendor

  The vendor has announced an agent layer without stating whether an assistant you bring can drive it. Recorded as unstated rather than guessed.

  None

2. Who holds the funds

3 of the 24 leave the funds in a wallet whose keys are yours. Most of the rest are ordinary custodial accounts, which is what a broker or an exchange normally is and is not a criticism of them. One product is neither, and it is this one: Fensory settles tokens to a wallet you control while perpetual futures margin sits at the venue. An earlier version of this page counted us among the non-custodial three. That was the only row here that did not match its own source, and it is worth saying plainly that it was the row that flattered us.

• 20

  The venue holds the funds

  The venue holds the funds. This is the ordinary arrangement for a broker or an exchange and it is what most of the field does.

  Alpaca MCP Server, Binance Agent OS, Bitpanda Fusion MCP, Bybit MCP Server, Coinbase for Agents, Composer MCP, cTrader AI Agent Connect, eToro Agent Portfolios, Gate MCP, Interactive Brokers AI Integration, Kraken CLI, OKX Agent Trade Kit, Phantom MCP, Robinhood Agentic Trading, Scalable Capital Agentic Investing, tastytrade MCP, TradeStation MCP, Trading 212, Walnut, Webull

• 3

  You hold the funds

  You hold the funds, in a wallet whose keys are yours. The agent is authorised to act within limits rather than handed the account.

  MetaMask Agent Wallet, Open-source Hyperliquid servers, Senpi

• 1

  Partly each, and the product says so

  Part with you and part at the venue, with the product saying so in its own description. One product is in this position and it is ours: tokens settle to a wallet you control while perpetual futures margin sits at the venue.

  Fensory

3. Whether the agent submits

The axis with the largest consequence and the least coverage. 3 of the 24 stop short of submitting and hand the order to a person.

• 20

  The agent can place the order

  The agent can cause a real order on a real venue without a person in the loop at the moment of submission.

  Alpaca MCP Server, Binance Agent OS, Bitpanda Fusion MCP, Bybit MCP Server, Coinbase for Agents, cTrader AI Agent Connect, eToro Agent Portfolios, Fensory, Gate MCP, Kraken CLI, MetaMask Agent Wallet, OKX Agent Trade Kit, Open-source Hyperliquid servers, Phantom MCP, Robinhood Agentic Trading, Senpi, tastytrade MCP, TradeStation MCP, Trading 212, Webull

• 3

  A person submits it

  The agent prepares the order and stops. A person submits it. This is a materially different product and the distinction is the one most often blurred in coverage of the category.

  Interactive Brokers AI Integration, Scalable Capital Agentic Investing, Walnut

• 1

  Withdrawn from sale

  Withdrawn from sale. Listed because directories and comparisons still recommend it, which is the reason to say so here.

  Composer MCP

How the category arrived

The four launches that made agentic trading a category rather than a handful of experiments, in the order they happened. The first product of any kind here dates to July 2025. Every dated launch on record, including the ones below, is on the research page.

A date marked “repository created” is when the public repository first existed, which is the earliest defensible public date. A date marked “reported” is one we have from coverage rather than from the vendor, and it is labelled rather than levelled up.

Every product, on all three axes

Alphabetical. Each classification is followed by the vendor’s own wording, so the call can be checked against what it came from rather than taken on trust. Fensory is in the list on the same terms as the rest, and nothing here is a recommendation.

What this page does not settle

Named rather than left out, because a reference page that implies completeness it does not have is worse than one that says where it stops.

• What stops a bad order

  The fourth question, and the one with the least comforting answer. It has its own analysis rather than a line here, because summarising it accurately takes more room than this page has.

  The control analysis

• Who can actually use each one

  Availability varies by country and many vendors do not state it. 9 of these products say nothing about where they are available at all. That figure has been wrong twice, both times because it was computed by pattern-matching the vendors' prose, and it is now a written classification with the evidence per product on the research page.

  What works outside the US

• How many people use any of this

  Nothing here counts users, volume or funded accounts. No vendor in the field publishes them and an estimate would be the least reliable thing on the page.

• Products we have not read

  Anything we have not opened the vendor’s own documentation for is absent rather than listed from a directory. The list grows when we have read the source, which means it lags announcements by design.

Questions people ask

What is agentic trading?

Agentic trading is when an AI assistant can cause a real order to be placed on a real trading venue, rather than only reading prices or suggesting what to do. The assistant reaches the venue through a published interface, most commonly an MCP server or a connector, and acts within whatever permissions the account gives it. The distinguishing feature is not that the software is clever. It is that a model can move money.

How is it different from algorithmic trading or a trading bot?

An algorithm or a bot follows rules a person wrote in advance, and it does the same thing every time the same conditions appear. An agent is given an instruction in language and decides for itself what to do about it, which means its behaviour is not fully enumerable before it runs. That is the whole difference, and it is why the question of what limits a bad order matters more here than it does for a bot.

Does an agent that only drafts an order count?

This page counts it and labels it separately, because the distinction is the one most often lost. A product where a person submits every order has a different risk profile from one where the agent submits, and several products described in the press as agentic trading are in fact draft-only. The comparison table on this page says which.

Who owns the funds while an agent is trading?

It depends on the product and the answer is not always one of the two obvious ones. At most venues the venue holds the funds, as it would for any brokerage or exchange account. At a few the funds stay in a wallet whose keys are yours and the agent holds a scoped authorisation instead. A small number sit between: the funds are at a third party the product does not control, or in a wallet created for the purpose whose signing runs as a service.

Is this list complete?

No, and it is dated rather than presented as current. It holds every product we have read on its own documentation, which is a smaller set than every product that exists. Products are added when we have read the vendor’s own page, never on the strength of a directory listing or another comparison.

Fensory is one of these products. Does that bias the page?

Fensory is in the table in alphabetical order, classified on the same three axes as everything else, and nothing on this page recommends it or any other product. The page exists because the definition is worth being the reference for, and a reference that argues for its publisher is not one. Where a judgement is ours rather than the vendor’s, the page says so.