MT5 AI Assistant: Features, Permissions and Limits

An assistant that can explain an Expert Advisor is useful. An assistant that can start that EA on a chart has a different kind of responsibility. Before using the MT5 AI Assistant, identify which of those tasks you are asking it to perform.

MetaQuotes has added native AI tools to MetaTrader 5 and MetaEditor. The latest release covered here is build 6230, with release notes dated 24 September 2026 and distribution announced for 25 September. Its new tools make the assistant more capable of preparing tests and managing programs on charts. They do not establish that a generated strategy is correct or profitable.

What is the native MT5 AI Assistant?

The native assistant combines a language model with tools that interact with the platform. Its usefulness depends on both parts: the model interprets a request, while available tools determine what it can actually read or change. A plausible answer in chat is not evidence that a tool ran successfully.

MetaQuotes introduced the broader agentic integration in build 6060. The company describes MCP as the connection between AI agents and platform functions in its architecture explanation. This can support sequences of platform operations, rather than text answers alone. To connect an agent outside the platform, such as Claude Code, to the same tools, see our MT5 MCP server setup and test.

The name matters when searching. A Market product, broker service or external application called an “MT5 AI assistant” is not necessarily the built-in MetaQuotes feature. Check the publisher, where the tool runs and which permissions it receives before comparing instructions or prices.

Terminal, MetaEditor and an EA do different jobs

An EA written with AI can remain completely rule based at runtime. Our MT5 EA built with AI compares prices with a time range and sends orders without making model requests. Conversely, an assistant with trading tools can affect a terminal without being a conventional EA. Calling both an “AI trading bot” hides the distinction you need to evaluate them.

What changed in build 6230?

The official build 6230 notes describe three relevant additions:

• Test preparation: retrieving an EA’s input parameters and preparing them for a Strategy Tester run.
• Chart programs: retrieving EA and script parameters, launching those programs on charts, removing EAs or indicators, and checking indicator state.
• Economic Calendar: retrieving countries, events, indicator values and records by ID.

These are documented capabilities, not results from our own assistant trial. The dated announcement schedules distribution for 25 September 2026. Your installed build and broker distribution can differ; check them directly.

For example, preparing a test configuration is useful only if you can inspect the chosen symbol, timeframe, dates, inputs and execution assumptions. An assistant can select a parameter correctly from a list and still choose a parameter that is inappropriate for the experiment you intended. That is a specification problem, not a compiler error.

Check availability in your installation

Start with the build number shown in your terminal’s About dialog. Record it alongside the date when reproducing an instruction. A screenshot from another build cannot confirm the features available in yours.

Then inspect the AI interface and provider settings in that installation. MetaQuotes’ build 6060 release notes, dated 23 July 2026, say that signing in with an MQL5.community account gives access to a free “MQL5 Lite” plan, with MQL5.community preselected as the AI provider. They also say you can use your own API keys for OpenAI, Anthropic, Gemini, DeepSeek, Ollama or other supported providers and choose the model in the terminal settings. That is MetaQuotes’ description, not something we observed. Verify the currently offered plan, applicable charges or limits and authentication requirements in the actual interface; this guide does not promise unlimited usage or that the free plan stays unchanged.

The paragraph above was corrected from a recheck of those release notes on 1 October 2026: they describe a free “MQL5 Lite” plan with MQL5.community as the default AI provider, not a default “MQL5 Lite” model as this guide previously said.

Keep those checks separate from your broker login. Access to MQL5.community, access to a model provider and authority to act on a trading account serve different purposes. A connected model is not evidence that trading is enabled, and a visible assistant is not evidence that every tool from the newest release is installed.

If a documented option is missing, compare your build with the relevant release notes first. Do not install an unrelated third-party plugin simply because its name resembles the missing native feature.

A read-only acceptance check you can reproduce

Before asking for a trading opinion, give the assistant an answerable technical task. Use a fresh test workspace, check the configured permissions, and supply the small CrossDirection function from our signal-logger project. Ask for the result of each input set without allowing file changes or terminal actions.

CrossDirection takes the fast and slow moving-average values of an older and a newer closed bar and returns 1 for an upward crossing, −1 for a downward crossing and 0 otherwise. The four input sets below are hand-selected numbers, not market data; the expected results are what the function returned in our native MQL5 test on 26 September 2026 in an isolated MT5 build 6230 terminal.

Record the installed build, provider/model shown, exact prompt, reply and any tool actions. Compare the reply with the source and these independently inspectable fixture results. If the assistant reverses the time order, calls the last case an upward crossing, or claims to have compiled without a compiler record, the response fails this task.

This is a prepared acceptance test, not an observed MT5 AI Assistant run. Our MQL5 fixture ran natively, but we have not completed an authenticated assistant conversation in the isolated installation. A correct table here must not be presented as proof that a particular model returned it.

If the assistant is missing or cannot answer

Do not use a model’s own assurance that it “has access” as the availability check. A successful read should identify an observable value that you can independently verify.

Understand the permissions before asking for actions

MetaQuotes’ build 6060 documentation describes trading-action controls that can allow, prohibit or require confirmation, with separate settings for network and command-line access. Inspect the actual controls in your build before using operational prompts. A sentence such as “only explain this” helps define the task, but it is not a substitute for the configured permissions. Source: build 6060.

For an explanation task, first use material that can be inspected without changing the account or environment. If you later request a change, name the target precisely: the file, chart, EA instance or tester configuration. “Clean up my terminal” leaves too much room for interpretation when tools can remove running programs.

Also distinguish preparing an operation from performing it. Loading an EA on a chart is already a change to the operating environment; it is not merely a recommendation in chat. Before authorizing that step, inspect its inputs and understand what its initialization and event handlers do.

Never paste trading passwords, API secrets or unredacted account details into an example prompt. When using a configured model provider, check that provider’s data handling and the context supplied to it. Do not assume that a tool being embedded in the terminal makes its model processing local.

Which tasks are a good start?

Use a task with a concrete output that you can compare against a source. The following matrix is a suggested workflow, not a list of operations we have executed with the assistant.

Prompt 1: explain a small function

Explain this MQL5 function without editing files or taking platform actions.
Identify the order of its time-series inputs and its equality conditions.
Walk through one upward crossing, one downward crossing and one no-cross case.
Separate what the code explicitly does from anything you are inferring.

Supply the function itself. If the answer calls its parameters “current” and “previous,” verify which closed bars those labels mean. Ambiguous wording here can reverse a signal later.

Prompt 2: describe inputs without choosing a strategy

Describe the available inputs of this teaching EA. Do not launch it or change
its inputs. State the unit and allowed range where these are documented.
Identify dependencies between inputs. If a label alone is insufficient to
explain behavior, say what source or documentation is missing.

An input named “Risk” is not enough to infer position sizing. It could mean a percentage of balance, a percentage of equity, a fixed amount, or something else. The useful answer points to the definition instead of inventing one.

Prompt 3: read an existing test report

Summarize this existing test report without starting another test or changing
settings. List the reported dates, symbol, inputs, modeling assumptions and
costs. Mark absent fields as unknown. Distinguish balance drawdown from equity
drawdown. Do not interpret a historical result as proof of future performance.

Compare every extracted number to the original. A fluent summary can still confuse the two drawdown measures or omit an open position. Our EA results checklist provides an independent framework for that review.

What still needs your review?

Check three things after an assistant response: the interpretation, the action and the outcome. Did it understand the rule? Did it use the intended tool on the intended target? Does the resulting file, configuration or platform state match the request?

A successful compilation answers a narrow question about code acceptance. A successful test run answers a different question about executing a particular experiment. Neither establishes a durable trading advantage. Keep historical backtests, historical out-of-sample checks, demo-forward observation and live records separate when evaluating any subsequent claim; our backtest vs live trading guide shows what each of the four describes and what it leaves unresolved.

For development, continue with the MT5 EA built from one prompt. For an EA already attached to a chart but doing nothing, use the diagnostic sequence to inspect initialization, permissions, signals and requests in order. An assistant becomes more useful when the evidence it needs is clear.