AI Assistant MCP in MT5 Beta Build 6194 has no file read tools

Built-in MCP tools strictly control access to internally permitted directories, and no external rules can change this. Even symbolic links within permitted directories are blocked to avoid any potential complaints.

The only correct way is to work inside the following directories:

catalog	rights
MQL5	read + write + execute
Common	read + write + execute
Tester	read
Logs	read

For temporary data, the agent uses the MQL5\Files\Temp directory.

Even in dangerous mode with shell commands, the agent (depending on the LLM model and provider) still tries to adhere to the restrictions on the system-declared directories.

Configuration - MCP and Launch - Welcome to Questions about the MQL5

Sergey Golubev 2026.10.11 04:28 #44

Forum on trading, automated trading systems, and trading strategy testing

New MetaTrader 5 Platform Build 6230: More AI Capabilities

Renat Fatkhullin , 2026.10.10 21:05

Since build 6253, subagents have been working in the agent.

----------------

Forum on trading, automated trading systems, and trading strategy testing

AI in MetaTrader 5: From Market Analysis to Trading Strategy Testing

Renat Fatkhullin , 2026.10.10 21:02

Since build 6253, the agent now uses subagents, improving efficiency and parallelism.

✓ subagent: delegate
✓ subagent: delegate
✓ subagent: metatrader5__get_trading_account_info
✓ subagent: metatrader5__list_directory

Both subagents are complete. Here's a demonstration in action.

What was used

Two parallel ad-hoc delegates via delegate :

delegate(instructions="...", async=true, max_turns=12)   → task 20261010_9
delegate(instructions="...", async=true, max_turns=12)   → task 20261010_10
load(source="20261010_9")     // дождаться 1-го
load(source="20261010_10")    // дождаться 2-го

Key principles of delegation:

Read-only exploration is freely parallelized —both subagents only read, and there are no resource conflicts.
async: true → tasks run in the background, then load(taskId) collects the results.
max_turns limits the cost of the subagent (here 12).
The instructions are self-sufficient —the delegate only knows instructions + source , so pre-flight and explicit rules (“do not trade,” “do not include login/balance”) are embedded directly into the task.

Sub-agent results (synthesis)

Subagent 20261010_9 — terminal facts (read-only):

chart_id	symbol	period	indicators / experts
134154936447362382	USDJPY	H1	0 / 0
23427280780490	EURUSD	H1	0 / 0
52113180007003	EURUSD	H1	0 / 0

Open charts : 3 , without indicators and experts.
Market Watch: The first 10 NASDAQ symbols (AACG, AADR, AAL, … ABEO). The full count was not returned by the tool.
Account: currency USD , margin mode netting ; experts_trade_allowed = false, mcp_trade_allowed = true.

Subagent 20261010_10 — MQL5 inventory (read-only):

Folder	.mq5 / .mqh / .ex5	Examples
MQL5\Experts	3 / 0 / 0	ChartInChart.mq5 , Server.mq5
MQL5\Indicators	20 / 0 / 6	bbands_stop_v1.mq5 , bbsqueeze.mq5 , correlation.mq5
MQL5\Scripts	28 / 0 / 10	1.mq5 , CanvasTest.mq5 , OpenCLTest.mq5 , OpenPosition1Lot.mq5
MQL5\Include	0 / 13 / 0	Benchmark.mqh , IsNewBar.mqh , OpenCL.mqh , PanelDialog.mqh
MQL5\Files	0 / 0 / 0	hello.py , My.db , ReportHistory-10556606.html , test.txt

When is what mode?

Mode	Call	When to apply
Sync	delegate(instructions=...)	One task, the result is needed immediately
Async	delegate(..., async=true) + load(id)	Several independent tasks in parallel
Source	delegate(source="name")	Ready recipe/agent
Combined	delegate(source="deploy", instructions="...")	Recipe + clarification

The rules I followed

Writes are file-based only : two delegates shouldn't touch the same file (conflict). Both delegates are read-only, so there was no risk.
Secrets are not output : the account subagent is explicitly prohibited from returning login/balance/equity.
Delegates are not coordinated : partition planning is done by the leading agent (me), and the results are synthesized here too.

Security note: One delegate noted a discrepancy—experts_trade_allowed = false on the terminal, while mcp_trade_allowed = true . Autotrading is currently disabled at the terminal level; this is an environmental issue; no trading activity was performed.