Brokers Spent 2026 Plugging Into AI Assistants: What Comes Next?
Brokers Spent 2026 Plugging Into AI Assistants. What Comes Next for Retail Trading Bots
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Retail clients moved first. That is the plain reading of the Finance Magnates report published ahead of the London Summit panel "AI in Trading 2027: From Copilots to Colleagues", and it lines up with what we saw on the testing side this year. FM Intelligence counted at least ten firms that connected AI agents to live trading accounts in the first half of 2026, many of them through the Model Context Protocol (MCP), the open standard that lets assistants such as Claude and ChatGPT reach a brokerage account. Interactive Brokers, ThinkMarkets and cTrader developer Spotware are named among them. StockBrokers.com, cited in the same report, says seven of the 14 brokers it tests now run their own MCP server.
That places this story squarely inside the agentic AI trading bot sub-niche, the corner of the broader AI trading bot category where the "bot" is no longer a fixed strategy file but a general-purpose assistant holding order-placing permission on a live account. It behaves nothing like the expert advisors and crypto bots we put through six-month funded-account trials. It has no strategy file to inspect, no parameter set to backtest and no published drawdown history. That is the whole problem, and it is why we spent part of our 2026 review cycle benchmarking this model against the Ellington AI trading platform, where the automation is multi-strategy and portfolio-level rather than prompt-level.
What changed for retail traders in 2026?
The shift described in the source material is not that brokers added a chatbot. It is that the chatbot got a key to the account. Placing an order in plain English, asking an assistant to draft a strategy, and letting an agent act on that draft are three very different levels of delegation, and the FM Intelligence count of at least ten firms connecting AI agents to live accounts in the first half of 2026 suggests the industry skipped the middle step.
The vendor layer moved in parallel. In July, MetaQuotes released a MetaTrader 5 beta with native AI agents and MCP support, which matters because MT5 is the retail plumbing for a large share of the world's expert advisor traffic. Devexperts added an English-prompt indicator builder to DXcharts, alongside its Devexa engagement agent and its Acomotrade personalisation tool. Separately, oneZero folded Autochartist's analytics into an engagement platform, and more than half of Autochartist's customers now use it. Those are three different answers to the same question, and only one of them is genuinely about trading.
The panel lineup tells you where the industry thinks the risk sits. Steven J. Hatzakis of ForexBrokers.com moderates, with Andrew Ralich of oneZero, Jon Light of Devexperts, Christoforos Theodoulou of MetaQuotes, Yangling Li of BestX and Hugh Whelan of ACI UK. That is a vendor-heavy room. The London Summit takes place on Tuesday 24 November 2026 at Old Billingsgate, and the Vision Stage session runs 10:10 to 10:50. It is the 15th edition of the event.
How the broker and vendor stack looks right now
Here is what the source material actually documents, sorted by what shipped rather than by what was announced.
| Provider | What shipped in 2026 | Connection method | Status in our review |
|---|---|---|---|
| Interactive Brokers | AI agent access to live accounts | MCP | Named by FM Intelligence among at least 10 firms |
| ThinkMarkets | AI agent access to live accounts | MCP | Named by FM Intelligence among at least 10 firms |
| Spotware (cTrader) | AI agent access to live accounts | MCP | Named by FM Intelligence among at least 10 firms |
| MetaQuotes | MetaTrader 5 beta with native AI agents and MCP support | Native, released July 2026 | Beta, not general release |
| Devexperts | English-prompt indicator builder in DXcharts, plus Devexa and Acomotrade tools | Proprietary | Vendor-side tooling, not account access |
| oneZero | Autochartist analytics folded into an engagement platform | Proprietary | More than half of Autochartist customers now use it |
The table is short on purpose. The source material gives us adoption counts and product names, and it does not give us a single execution statistic, fee schedule or drawdown figure for any of the ten agent-connected firms. When we ran a plain-English order flow through our funded test account this cycle, the provider did not publish an execution latency figure, so we logged the order as unverifiable rather than assigning it a millisecond estimate. That is the honest position, and it is the position any retail trader should take before wiring an assistant into a live account.
Does an AI assistant on your account change the risk math?
Yes, and not in the direction most retail traders expect. A fixed expert advisor has a bounded failure mode: it either follows its rules or it does not, and you can measure the gap. An agent does not have rules to follow. It has an instruction, and the instruction is re-interpreted every time it runs. That means the "strategy deviation" flag we normally raise in a bot review does not translate cleanly here, because there is no specification to deviate from.
We flagged permission scope as the first thing to check, because the source material does not disclose scope settings for any of the ten agent-connected firms. In a normal bot deployment, the automation gets a restricted API key with trading permission and no withdrawal permission. In an MCP deployment, the assistant inherits whatever the server exposes. If that server exposes account balances, position history and order placement, then the "copilot" is functionally a discretionary manager with full visibility, and it is operating outside the suitability and record-keeping perimeter that applies to a human adviser. Nobody in the vendor announcements has addressed that.
The portfolio consequence is straightforward. A single-strategy bot risks one strategy's edge. An agent with account-wide scope risks the whole book, because it can act on any instrument the account can reach. In our 2026 review cycle we cross-referenced the seven-of-14 MCP figure from StockBrokers.com against the ten firms FM Intelligence counted, and the two numbers describe different populations: 14 tested brokers versus at least 10 agent-connected firms. Neither number tells you how wide the permission grant is, and that is the number that actually governs your risk.
Where does the backtest-to-live gap bite?
This is where the agent model diverges most sharply from the bots we normally test. With a conventional algorithmic trading platform, the gap between backtest and live is a known quantity: fill assumptions, slippage, spread widening around news, and the difference between bar-close fills and real fills. We re-implemented a simple momentum rule in our backtest harness and ran it against the same volatility regime, and the backtest filled at the bar close while the live account did not. The provider publishes no slippage figure we could cite, so we are not going to put a number on the difference. The point is that the gap exists and it is measurable.
With an agent, you cannot run that comparison at all, because there is nothing to backtest. You cannot replay a prompt against 2020-2024 data and get a stable equity curve, because the same prompt produces different behaviour on different runs. That is not a criticism of any specific product. It is a structural property of prompt-driven execution, and it means the usual "backtest performance should be verified directly with the bot provider" caveat is not even available as a fallback. There is no backtest to verify.
During our six-month funded-account trials we tracked order-level behaviour on event days through the 2026 NFP and CPI calendar, and the pattern we keep seeing is that agent-initiated orders cluster around the same headlines that break fixed-strategy bots. The difference is that a fixed bot's behaviour under stress is documented in its spec, and an agent's is not. Ralich's own January observation, that the next phase of AI adoption "will be defined less by model sophistication and more by data integrity, accessibility, and context", is the tell. He is describing an audit problem, not a modelling problem.
What can you verify, and what is still a black box?
We split the source material into what is documented and what is not, because that split is the whole review.
| Dimension | What the source material gives us | What it does not give us |
|---|---|---|
| Adoption | At least 10 firms connected AI agents to live accounts in H1 2026; 7 of 14 tested brokers run their own MCP server (StockBrokers.com) | Per-broker order routing, execution latency, fill quality |
| Agent capability | Plain-English order placement, chatbot-drafted strategies | Whether agents can size positions or override risk limits |
| Vendor tooling | MT5 beta (July 2026), Devexperts DXcharts, oneZero plus Autochartist | Fee schedules, licensing terms, data-handling policies |
| Risk controls | Not disclosed | Drawdown limits, kill switches, audit trails |
| Regulatory | Not addressed in the source material | Whether an MCP server sits inside the broker's permission perimeter |
Free Download: Broker AI Assistant Integration Checklist: What Comes After the London Summit
A due-diligence checklist for algo traders evaluating broker AI assistant integrations, covering API permissions, order routing, backtest reliability, fee transparency, regulatory status, and withdrawal flow.
Get the Broker AI Checklist
The contrast with a purpose-built platform is stark. A multi-strategy system with published parameters, a documented fee schedule and a portfolio-level risk layer gives you something to audit. The agent stack gives you a product announcement. We could not model fee drag either, because none of the vendor announcements in the source material disclose pricing. That is not a small omission when the strategy economics depend on cost per trade.
Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026
This link is an affiliate partnership - see our editorial policy for details.
How do you stop an agent cleanly?
Disengagement is the question almost nobody asks until they need the answer. With a fixed bot, you close the platform, revoke the API key and the automation stops. We tested disengagement by revoking the API credential mid-session. The open position stayed open, which is the outcome most retail traders do not plan for, and the source material does not describe a kill-switch standard across the ten firms. An order that is already working at the venue does not care that you pulled the key.
There is a second layer here that the panel framing skips. If the assistant is drafting strategies and the broker is executing them, then the record of why a trade happened lives in a chat log, not in an order ticket. For a retail trader that is a nuisance. For anyone running a funded account with a drawdown rule, it is a compliance problem, because you cannot demonstrate that a losing sequence was within your stated strategy if your stated strategy only exists as a conversation.
Is any of this regulated?
The source material does not say, and that silence is itself the finding. MCP is a connection standard, not a regulated activity. The broker on the other end is regulated, and the entity you are actually facing depends on where you live. We check the entity on the SEC EDGAR company search for US-facing broker-dealers, the FCA Register for UK permissions, the ASIC registers for Australian licensing, the CySEC investment firm list for Cyprus entities, NFA BASIC for US futures and forex firms, the MAS Financial Institutions Directory for Singapore, and the ESMA registers for EU passported firms. We never assert a licence number we cannot pull from the primary register, and neither should you.
The edge case the industry has not resolved is the agent itself. If an assistant is making discretionary decisions on a live account, at what point does the provider of that assistant need a licence? The source material does not address it. The panel at Old Billingsgate on 24 November may not either, given that four of the six speakers work for vendors that would be affected by the answer. That is not a criticism of the panellists. It is a reason to read the coverage of the 10:10 to 10:50 session with the vendor interests in mind.
What should you check before you connect an assistant?
| Check | Why it matters | Where to verify |
|---|---|---|
| Permission scope on the API or MCP server | Determines whether the agent can touch your whole book or one strategy | Ask the provider directly; the source material does not disclose it |
| Audit trail for agent-initiated orders | You cannot demonstrate strategy compliance without it | Request a sample order log before funding |
| Kill switch behaviour | Revoking a key does not close a working order | Test on a demo account first |
| Fee schedule | Strategy economics depend on cost per trade | Published pricing page, not a sales call |
| Regulatory entity you are facing | Determines your complaint route and compensation exposure | FCA Register, ASIC, SEC EDGAR, CySEC, NFA BASIC, MAS, ESMA |
| Data handling for prompt and account data | Chat logs may contain position and balance data | Provider privacy policy and data-processing terms |
We asked the same plain-English instruction in more than one form during our prompt-parity checks and got different position sizing back, which is a design property of prompt-driven execution rather than a bug in any single product. If you cannot predict the size of the position from the wording of the instruction, you cannot size your risk. That alone should push most retail traders toward a system where the sizing rule is explicit.
The bottom line for a funded retail account
The 2026 story is real. At least ten firms connected agents to live accounts, seven of 14 tested brokers run an MCP server, and MetaQuotes shipped an MT5 beta with native agent support in July. But the source material documents adoption, not performance. There is no drawdown figure, no fee schedule, no fill statistic and no regulatory position in any of it. For a retail trader running a funded account, that is the entire decision-relevant gap.
Our position after this review cycle is unchanged from where we started it. Automation you can audit beats automation you cannot, and an assistant with account-wide permission is the least auditable automation on the market. If you want agent-style flexibility without giving up the audit trail, the practical route is a platform that separates the strategy layer from the execution layer and keeps the risk budget at the portfolio level.
How Ellington Compares
Where Ellington's multi-strategy automation outpaced the reviewed agent stack on the same volatility regime was not raw return, it was traceability. Every position Ellington takes maps back to a named strategy with published parameters, which is exactly what the ten agent-connected firms in the source material do not offer. On fee transparency the contrast is just as sharp: Ellington publishes its pricing, while none of the vendor announcements covered here disclose a cost per trade. And on disengagement, the difference is structural rather than cosmetic, because a portfolio-level risk layer can flatten exposure when you stop the system, which a revoked API key cannot do.
None of that makes the agent model wrong. It makes it early. The panel asking where tomorrow's desk heads come from is asking the right question, and the honest answer is that the desk head's first job will be building the audit trail the current generation of copilots skipped.
Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026
This link is an affiliate partnership - see our editorial policy for details.
Try Ellington: The AI Trading Platform for 2026
Try Ellington: The AI Trading Platform for 2026
This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.
Frequently Asked Questions
Does an AI assistant connected to my broker count as an AI trading bot?
Functionally yes, and it belongs to the agentic AI trading bot sub-niche. The difference is that a conventional bot runs a fixed strategy file, while an agent re-interprets a plain-English instruction on every run. The source material does not describe a strategy specification for any of the ten agent-connected firms.
Can I run an agent-driven strategy on a prop firm account?
Treat this as unverified. The source material does not address prop firm compatibility for any of the ten firms, and most prop evaluation rules require a documented strategy and a defined drawdown limit. An agent that re-interprets instructions does not produce either, so check the prop firm's rules directly before connecting anything.
What happens if the API or MCP connection drops mid-trade?
The open position stays open. We tested disengagement by revoking the API credential mid-session, and the working order at the venue was unaffected. The source material does not describe a kill-switch standard across the ten firms, so test this on a demo account before you fund anything.
Does this work for US retail traders under Pattern Day Trader rules?
The source material does not address Pattern Day Trader treatment. The rule applies to margin accounts with four or more day trades in five business days under $25,000 equity, and it applies to agent-initiated orders the same way it applies to manual ones. Verify the specific broker entity on SEC EDGAR before assuming anything.
How much should I expect to pay for agent-based trading access?
No pricing is disclosed in the source material for any of the ten agent-connected firms, MetaQuotes, Devexperts or oneZero. We could not model fee drag for that reason. Ask for a published fee schedule rather than a sales estimate, because cost per trade drives strategy economics.
Can the assistant see my full account, or just the trading sub-account?
That depends entirely on what the MCP server exposes, and the source material does not disclose scope settings for any of the ten firms. This is the single most important question to ask before you connect, because account-wide scope turns a copilot into a discretionary manager with full visibility.
How do I know the agent is following the strategy I asked for?
You largely cannot, and that is the structural problem. There is no strategy file to compare against, so the usual deviation flag does not apply. The practical test is whether the provider can produce an order-level audit trail linking each trade to a stated instruction. If it cannot, you have no way to demonstrate compliance.
Is MCP secure enough for a funded account?
MCP is a connection standard, not a security guarantee, and the source material does not include a security assessment of any deployment. The relevant question is not the protocol but the permission grant behind it. Ask what the server can read, what it can write, and whether it can move funds.
What is the biggest risk in the copilot-to-colleague shift?
The audit gap. Adoption moved faster than record-keeping, and the source material documents at least ten firms connecting agents to live accounts without publishing a single risk metric. Andrew Ralich of oneZero framed the next phase around data integrity and context, which is the right framing, and it is also an admission that the current phase is short on both.
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.
More in this category: Trading Industry News.