State of Agentic Trading, Q3 2026

The headline figures

Counted on 2026-10-18. Each is computed from the product table rather than written down, so this page and the comparisons it came from cannot disagree.

What stops a bad order

The question every product in this category answers differently, and the one most of them do not answer at all. A limit written into a prompt is a request the model can be argued out of. A limit enforced on the account is a rule. These are the shapes that exist, counted from what each vendor documents rather than from what it claims.

• 14

  Nothing documented. The limit is how much money is in the account

  Alpaca MCP Server, Binance Agent OS, Bitget GetClaw, Coinbase for Agents, Composer MCP, Interactive Brokers AI Integration, Open-source Hyperliquid servers, Phantom MCP, Robinhood Agentic Trading, Scalable Capital Agentic Investing, TradeStation MCP, Trading 212, cTrader AI Agent Connect, eToro Agent Portfolios

• 4

  A flag, scope or key that decides whether trading is reachable at all

  Bitpanda Fusion MCP, Bybit MCP Server, Gate MCP, OKX Agent Trade Kit

• 1

  A person approves each order. Real, and it stops being a control the moment nobody is watching

  Walnut

• 2

  A ceiling on the size of any single order

  Fensory, Webull

• 2

  A ceiling that bounds a period rather than an order, whether on losses or on spend

  MetaMask Agent Wallet, Senpi

• 2

  The software itself refuses to act without a confirmation step, rather than asking the model to

  Kraken CLI, tastytrade MCP

A capability gate is not a cap. It decides whether trading is reachable at all and says nothing about how much can go through once it is. Counting the two together would make this field look far safer than it is.

The category, quarter by quarter

Every dated launch

Oldest first. A date marked “repository created” is when the public repository first existed, which is the earliest defensible public date and usually sits a little before the announcement.

Who can actually use these

Availability as each product’s own page states it, which is not the same as where its parent company operates. 10 of the 25 do not state it at all, which is the least reported thing about this category. A product being available where you live is not the same as a product saying so, and an agent cannot read an omission.

What this dataset does not have

Named rather than omitted. A dataset that hides its gaps is not a dataset, and two of these are the questions a reader most wants answered.

• Fensory ledger statistics: tool calls, agents connected, asset mix, client mix, repeat rate

  Not yet confirmed internally. A company publishing research about its own category does not get to quote unverified figures about itself.

• Total on-chain value of tokenized equities

  No source we control, and repeating another aggregate without reading its methodology is how a figure like this goes wrong.

• Adoption: how many people actually use any of this

  Nobody publishes it, ourselves included. Counting products is not counting users, and this page does not pretend otherwise.

Questions people ask

What counts as an agentic trading product here?

One where an AI assistant can cause a real order to be placed on a real venue. A product that only reads prices or drafts an order for a human to submit is not counted, and where a product does both we count the execution path. Market-data servers are excluded entirely.

Where do these figures come from?

Each product was read on its own documentation, its own repository, or its own terms, never on a directory listing or another comparison. Repository figures come from the GitHub API. Every product links to the source its claims were read on.

Is this complete?

No, and the page says which parts are missing rather than implying coverage. A re-sweep in September added four official exchange servers an earlier pass had missed, several more are named as reported but unverified, and nothing here counts users rather than products.

Why are products that only draft an order counted here at all?

Because leaving them out would hide the distinction rather than report it. Four of the products here do not place an order on their own: three prepare one for a person to submit and one has been withdrawn from sale. All four are described as agentic trading elsewhere, so a reader comparing them deserves to see which is which rather than a list that quietly excludes them. The counts on this page say which of the two groups each figure refers to.

Fensory publishes this and sells a trading account. Why trust it?

Because the figures are checkable and the page does not flatter us. Fensory is one of two products documenting a cap on a single order, and a per-order cap is the narrowest protection here: it bounds one mistake and does nothing about thirty in a row, which is the failure an unattended agent actually has. Senpi covers that and we do not. Both facts are in the counts below.