mcp-mt5
Model Context Protocol server for the MetaTrader 4/5 build pipeline, enabling LLM-driven compilation, deployment, backtesting, and log tailing without UI interaction.
README
mcp-mt5
Model Context Protocol server for the MetaTrader 4/5 build pipeline. Compile MQL sources, deploy compiled EAs, run Strategy Tester, parse reports, tail logs — all driven by an LLM agent without touching the MetaTrader UI.
What this is — and what it isn't
| ✅ This server | ❌ Not this server |
|---|---|
| MetaTrader dev harness — compile, deploy, backtest, parse | Live trading (orders, positions, quotes) |
Wraps MetaEditor64.exe / terminal64.exe CLI directly |
Wraps the MetaTrader5 Python package |
| Runs entirely offline against installed terminal | Connects to a broker server |
| Iterates strategies before they go live | Executes strategies in production |
Use case: an LLM agent edits
.mq5source → compiles → deploys → runs Strategy Tester → reads report → adjusts → repeats. No broker login, no human in the loop, no risk of real-money execution.
For runtime trading, pair this with a live-trading MCP — they target different layers and compose well.
Tools
The server exposes 32 tools and 3 MCP resources across nine categories.
🔍 Discovery & terminal selection
| Tool | Description |
|---|---|
env_info |
Dump resolved paths, terminal hash, edition, and missing-component issues |
list_terminals |
Enumerate every MT4/5 terminal data folder under %APPDATA%\MetaQuotes\Terminal along with each origin.txt install path |
select_terminal |
Switch the active terminal data folder mid-session by origin path, hash, or install dir — handy for testing across multiple brokers |
🔨 Build & deploy
| Tool | Description |
|---|---|
compile |
Invoke MetaEditor CLI on a .mq4/.mq5/.mqh source. Returns structured errors[]/warnings[] (file, line, column, error code, message) plus log excerpt |
compile_and_deploy |
Compile, then copy the resulting .ex4/.ex5 into the terminal's Experts/ folder in one call |
syntax_check |
Same as compile but uses MetaEditor's /s syntax-only mode for faster feedback |
smoke_test |
Compile + deploy + run a 1-day headless backtest + scan the journal for runtime errors. Catches problems that pass compile but fail at runtime |
deploy_ea |
Copy a compiled binary into Experts/ (with optional rename) |
install_include |
Copy a .mqh header into the terminal Include/ folder — handy for libraries like LiveLog.mqh |
list_experts |
Enumerate Experts/ recursively with size and modification time |
🔎 Source analysis
| Tool | Description |
|---|---|
extract_inputs |
Parse input <type> <name> = <default>; declarations into structured records |
gen_tester_inputs |
Auto-build a [TesterInputs] block from EA source (translates PERIOD_* enums to numeric codes), optionally write into an existing tester.ini |
resolve_includes |
Recursive #include resolution that reports missing files and circular references |
find_symbol |
Grep a symbol across MQL files, skipping comments and string literals |
code_metrics |
LOC, function count, max nesting per file — or aggregated across an entire tree |
extract_doc |
Pull MetaEditor //+--+ //| ... +--+ doc blocks out as markdown |
find_magic_collision |
Detect duplicate magic-number assignments across the project |
⚠️ Lint & validation
| Tool | Description |
|---|---|
lint_basic |
Structural rules: missing OnInit/OnDeinit, unused inputs, hardcoded magic numbers, hardcoded symbol literals |
check_deprecated |
Flag MT4-style API calls (OrderSend, Ask, AccountBalance, …) with CTrade/MT5-API replacement suggestions |
validate_tester_ini |
Sanity-check a tester.ini (required keys, date format, numeric ranges) and cross-check [TesterInputs] against the EA source declarations |
🎨 Format
| Tool | Description |
|---|---|
format_mql |
Format a source file via clang-format (treats MQL as C++ with an MQL-friendly default style) |
format_check |
Same as above but reports whether changes are needed without writing the file |
✏️ Refactor
| Tool | Description |
|---|---|
rename_symbol |
Whole-word rename across all MQL files in a tree, with dry_run preview |
extract_function |
Brace-aware extraction of a contiguous block into a new helper function — inline or into an external .mqh |
📊 Strategy Tester
| Tool | Description |
|---|---|
patch_tester_ini |
Programmatically update keys in a tester.ini (e.g. Tester.Symbol, Tester.FromDate, TesterInputs.RiskPct) before running |
run_backtest |
Launch terminal64.exe /config:tester.ini, optionally headless (when ShutdownTerminal=1), and return the latest tester log path |
parse_optimization |
Best-effort parser for the latest .opt (optimization passes) binary file |
top_passes |
Sort optimization passes by a chosen criterion and return the top N |
read_tester_report |
Locate and parse the latest tester HTML report into a structured summary (net profit, profit factor, drawdown, trade counts, etc.) plus a sample of trade rows |
compare_reports |
Diff two tester reports key-by-key with absolute and percent deltas |
regression_check |
Verify a candidate report stays within guard thresholds vs a baseline (e.g. "net_profit may not drop more than 5%") |
kill_terminal |
taskkill if the terminal hangs |
📝 Logs & snapshots
| Tool | Description |
|---|---|
tail_log |
Tail the last N lines of either Files/LiveLog.txt, the daily Logs/YYYYMMDD.log, or the most recent tester log. Optional structured parse into {ts, source, message} records |
snapshot_sources |
Freeze a copy of source files into a timestamped folder with a manifest.json |
list_snapshots |
Enumerate previously captured snapshots |
📡 MCP resources
Live, re-readable URIs that an MCP client can poll instead of calling a tool repeatedly.
| URI | Description |
|---|---|
mt5://livelog |
Latest tail of MQL5/Files/LiveLog.txt |
mt5://journal |
Today's daily MT5 journal log |
mt5://tester-log |
Most recent Strategy Tester journal |
Quick start
Install
pip install mcp-mt5
Requires Windows + an installed MetaTrader 4 or 5 terminal.
Register with an MCP client
Most MCP clients accept a JSON entry under mcpServers. The server inherits its configuration from environment variables:
{
"mcpServers": {
"mt5": {
"command": "mcp-mt5",
"env": {
"MT5_INSTALL": "C:\\Program Files\\MetaTrader 5"
}
}
}
}
Refer to your client's documentation for the exact config file location.
Verify the install
Once registered, ask your agent to call env_info:
{
"edition": "mt5",
"install": "C:\\Program Files\\MetaTrader 5",
"terminal_hash": "<32-char-hex-hash>",
"metaeditor": "C:\\Program Files\\MetaTrader 5\\MetaEditor64.exe",
"experts_dir": "C:\\Users\\<you>\\AppData\\Roaming\\MetaQuotes\\Terminal\\<hash>\\MQL5\\Experts",
"issues": []
}
An empty issues array means everything is wired up correctly.
Configuration
Resolution priority for the MetaTrader install + data folder:
- Explicit env vars (below)
- Auto-scan of
%APPDATA%\MetaQuotes\Terminal\*\origin.txtfor a folder whose origin matchesMT5_INSTALL - Portable mode fallback (data colocated with install dir)
| Env var | Default | Notes |
|---|---|---|
MT5_INSTALL |
C:\Program Files\MetaTrader 5 |
Install dir containing terminal64.exe |
MT5_DATA |
(auto-detected) | %APPDATA%\MetaQuotes\Terminal\<hash> |
MT5_TERMINAL_HASH |
(auto-detected) | 32-char folder name |
MT5_EDITION |
mt5 |
Set to mt4 for MetaTrader 4 |
MT4 support
Set MT5_EDITION=mt4 and point MT5_INSTALL at your MT4 install. The server switches to metaeditor.exe (32-bit), terminal.exe, and the MQL4/ data tree automatically.
Example workflow
A typical LLM-driven iteration loop:
1. env_info → verify paths
2. compile_and_deploy source="MyEA.mq5" → 0 errors, .ex5 deployed ✅
3. patch_tester_ini config="tester.ini" updates={
"Tester.Symbol": "EURUSD",
"Tester.FromDate": "2025.01.01",
"TesterInputs.RiskPct": "1.5"
}
4. run_backtest config="tester.ini" wait=true
5. read_tester_report → summary.net_profit = 1234.56
summary.profit_factor = 1.45
6. tail_log mode="tester" lines=200 structured=true → diagnose journal warnings
7. <edit Signal.mqh based on findings>
8. → loop back to step 2
A sample tester.ini
; Launch: terminal64.exe /config:tester.ini
; Period codes: M1=1, M5=5, M15=15, H1=16385, H4=16388, D1=16408
; Model: 0=Every tick, 1=1 min OHLC, 4=Real ticks
[Tester]
Expert=MyEA
Symbol=EURUSD
Period=M15
Model=1
FromDate=2024.01.01
ToDate=2024.12.31
Deposit=10000
Currency=USD
Leverage=500
Visual=0
ShutdownTerminal=1 ; required so run_backtest can wait for the run to finish
Report=tester_report
[TesterInputs]
; ParamName=value||start||step||stop||(N=fixed|Y=optimize)
; RiskPct=1.0||0.1||0.1||3.0||N
A more complete sample lives at examples/tester.ini.
Development
git clone https://github.com/PHUICMT/mcp-mt5
cd mcp-mt5
pip install -e ".[dev]"
pytest # runs the 18-test suite
ruff check src tests # lints
CI runs on Windows for Python 3.10, 3.11, and 3.12 against every push to main. Tagging a release (e.g. v0.2.0) triggers an OIDC publish to PyPI.
Project layout
mcp-mt5/
├── src/mcp_mt5/
│ ├── server.py # FastMCP tool definitions
│ ├── paths.py # Layout detection + origin.txt scan
│ └── parsers.py # Compile log + tester HTML report parsers
├── tests/ # 18 pytest tests, no live MT5 required
├── examples/ # Sample tester.ini + client config
└── .github/workflows/ # CI + PyPI release
Limitations
- Windows-only. MetaTrader CLI binaries don't ship for Linux/macOS. Wine ports may work but are untested.
- No live broker access. This server intentionally never authenticates to a broker. Use a separate MCP server for runtime trading.
- Tester report parsing is best-effort. MetaTrader's HTML output isn't a stable schema; the raw HTML is also returned alongside the parsed structure so you can fall back to text inspection when needed.
- Optimization runs are not parsed yet. Single-pass backtests are fully supported;
.optresults are on the roadmap.
Roadmap
All v0.3.x roadmap items shipped in v0.4.0. Future ideas:
- Real tree-sitter MQL grammar for
extract_function(current implementation is brace-counting + regex) - WebSocket transport for long-lived sessions (currently stdio only)
- Linux/Wine port for non-Windows agents
License
MIT © 2026 PHUICMT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。