Retail FX and CFD platforms have added ways for AI agents to place and manage trades on live customer accounts over the past year, giving dealing desks a new visibility problem. MetaQuotes, Spotware and Match-Trade have each introduced MCP-based connections, but orders from an agent can reach a broker through the same route as a customer's own platform activity.
That makes it harder to use past behaviour to profile a client and decide how to handle their flow. Brokers have long dealt with automated strategies, including Expert Advisors (EAs), but an AI agent may not carry a distinct identifier that shows who or what initiated an order.
AI agents gain live account access
The Model Context Protocol, or MCP, sets a common way for AI agents to connect with software. A server built for a trading platform or by a third party makes available a set of actions. A trader could, for example, ask an agent such as Claude to close a position, and the agent can send that instruction to the account.
MetaQuotes added MCP support within MetaTrader 5, Spotware introduced cTrader AI Agent Connect, and Match-Trade first made an MCP connection available to brokers before extending it to traders. With the official platform connections, agents can carry out many of the actions available to a trader using the platform directly.
How access is granted depends on the connection. A remote service may ask the trader to provide an account token to an AI tool; a local arrangement connects the agent through the trading platform on the trader's computer. The user can generally decide whether the agent can trade or must request approval. If trading is enabled without approval, the agent can open, change and close positions. Some official options also allow read-only access or confirmation before an order is sent.
Other routes include open-source servers that use platform APIs, hosted services designed to work with multiple platforms, and servers built by traders with coding skills. Their safeguards vary. Some may not provide the approval settings available in official connections. With a hosted service, account credentials or a token may be held by the service provider. A broker may have no record of the agent's instructions or actions unless it operates the server itself or the platform provider supplies that information.
Visibility varies across broker setups
The additions have appeared over the past year, but the reporting does not give individual launch dates. The visibility issue applies differently depending on the broker's platform and how a customer connects an agent. ThinkMarkets, for example, operates its own MCP server and can view the agent's activity. The reporting says that level of access depends on the broker controlling the server; firms using an external platform rely on its provider for similar records.
MetaTrader can show a dealing desk whether an order came from desktop, mobile, web or an EA. That information helps dealers identify automated activity. But AI orders do not have their own label. Some open-source MT5 connections require algorithmic trading to be enabled, so their orders may appear as EA trades instead. The report also describes cases where EA activity has been presented to dealer systems as mobile activity.
Trading patterns can offer clues, though they do not provide proof of AI use. Human customers may trade at irregular intervals, change course under psychological pressure or focus on active market hours. EAs tend to follow the conditions encoded in their logic, often using technical rather than fundamental inputs. Orders arriving within milliseconds can indicate automation. AI agents, by contrast, may combine different inputs and produce less consistent patterns. A server's records can show which account an agent accessed and what it did, but most brokers cannot see those logs unless they control the server or receive the data from the platform provider.
A report by Finance Magnates also describes the limits of existing dealer checks, drawing on experience profiling clients at major brokers. The report says that identifying EAs can already be difficult when their activity is not labelled as expected; agent trading adds another source of uncertainty.
What this means on the desk
Client profiles used for A-book or B-book allocation may stop reflecting how an account is trading if a customer changes from manual activity to an agent. Shared models, prompts or open-source connections could also cause multiple clients to respond similarly to market conditions, increasing the chance of concentrated positions. Dealers may need to reassess profiles when behaviour changes and watch for flow that moves together.
There are also execution and record-keeping questions. An agent can misunderstand an instruction, and a broker without the server's logs may not be able to establish what the customer asked it to do. An agent connected to faulty prices, calendars or other third-party tools may place orders based on bad inputs. If credentials or tokens are compromised, another party could trade through the account in a way that appears to come from the customer.
Operators should check which connection methods their clients can use, where tokens and login details are stored, what approval controls are available, and whether agent actions are logged. They can also ask platform providers whether MCP-originated orders can carry a distinct label and whether server records can be made available to the broker. Until that visibility exists, dealing desks must treat changes in client behaviour and correlated activity as signals to investigate, not proof of who or what placed an order.
Platforms · Software · Providers · Markets