Trading platforms add AI agents, but brokers cannot identify orders

A dealing desk monitor displaying trading charts, illustrating the challenge of identifying AI-generated orders.
Photo by AlphaTradeZone on Pexels

Over the past year, MetaQuotes, Spotware and Match-Trade have enabled AI agents to interact with retail FX and CFD accounts through their platforms. For dealing desks, the operational gap is identification: most brokers have no dependable signal showing whether an order came from an AI agent or from the client.

The connections use the Model Context Protocol (MCP), which gives compatible AI tools a standard way to call software functions. A trader can link an agent such as Claude to a trading account and ask it to take an action.

Depending on the server and its permissions, those actions can include opening, changing or closing positions. Some configurations require approval before a trade is sent; others let the agent act without it.

Remote connections can use an account token entered into an AI tool. A local configuration instead works through the trading platform on the client's computer. In the latter case, AI instructions may travel through the same session as the trader's own orders.

Existing platform labels can distinguish some sources, such as desktop, mobile, web or an Expert Advisor, but the source says there is no dedicated AI marker. That complicates profiling and A-book versus B-book decisions, particularly if a client's behaviour changes or several accounts respond similarly to market conditions.

Platform access outpaces dealer visibility

MeasureReported detail
Rollout periodOver the past year
Automated order speedWithin milliseconds
Publication date12 November 2026
Adoption volumeNot reported in the source

The identification problem is not simply a matter of adding a new label to an order record. An AI agent using a local connection may appear to the broker as activity from the trader's usual session. Some open-source MT5 MCP bridges require algorithmic trading to be enabled, which can make orders appear under an EA label instead. Neither route necessarily tells the desk that an AI made the decision.

Trading patterns may offer clues, but they do not establish what software generated an order. Human activity can be irregular; EAs tend to follow coded conditions; agents can combine different inputs and produce less consistent patterns.

The source also notes that dealers have encountered ways for EA orders to appear under another platform source. That makes behaviour-based detection an imperfect substitute for an explicit platform signal.

Server ownership determines audit access

MetaQuotes added native MCP support to MetaTrader 5, Spotware introduced cTrader AI Agent Connect, and Match-Trade first made an MCP server available to brokers before extending it to traders. Official servers offer controls including read-only access and per-trade approval, according to the source.

Other connections include community-built software, hosted services spanning multiple platforms, and servers coded by traders themselves. Their safeguards depend on their implementation.

ThinkMarkets has built its own MCP server, giving the broker visibility into the agent's actions, the source reports. That access is not generally available to brokers using platforms they do not operate. For other connections, orders reach the platform through its usual API or terminal, and the broker may not receive the MCP server's records of which account the agent used or what it did. The story was first reported by Finance Magnates.

The missing record can matter when an agent misreads an instruction, uses unreliable information from an external tool or acts on a compromised token. The broker may see only a normal order, while the client may dispute the result. Without a server log, the broker could lack evidence of the instruction and the agent's actions. The source also flags possible complaints, reputational harm, regulatory scrutiny and unintended suspicion of market manipulation.

Where this lands in the stack

For dealing, risk and operations teams, the immediate checks are about observability and control. Ask platform providers whether MCP-originated orders can carry a distinct identifier, and whether broker systems can retain the server's account and action records. Confirm which execution paths make an order look like a desktop, mobile, web or EA instruction.

Review how client profiles are refreshed when trading behaviour changes, and how correlated activity is monitored across accounts. For hosted or trader-built servers, establish where credentials or tokens are held and what approval and access controls exist. The source does not quantify adoption, so operators should also ask what evidence they can use to track use of these connections across their own client base.