junet.storeAI × MQL5 × Open Source
VIChuyển sang Tiếng Việt
AI agents

Connect MQL5 CodeGraph MCP to Codex and other AI agents

Understand the local MCP boundary, register mql5-codegraph-mcp per client, select a separate trusted project root, and maintain honest RAM-only snapshots.

2026-08-022 min read
MCPCodexAI agents

MCP lets an agent call analysis tools through an explicit contract instead of guessing commands and interpreting arbitrary JSON. A GitHub URL, however, does not turn every agent into a prepared local environment.

Minimum prerequisites

The client needs local terminal and filesystem access, Internet access for wheel installation, Python 3.11+, package-install permission, and a way to register a stdio MCP server. After installing v0.3.0, the executable is:

mql5-codegraph-mcp

Cursor, Claude Code, and Antigravity use their own MCP configuration formats. Follow the current client documentation to register this executable; do not blindly copy one configuration file between applications.

A working session

  1. Start a fresh task inside the real MQL5 repository.
  2. Call project_status to see whether the server has a snapshot.
  3. Index only a local project root selected and trusted by the operator.
  4. Use query, architecture, context, or impact before proposing code changes.
  5. Re-index after source changes or an MCP process restart.

The experimental Codex plugin in the repository supplies five workflow skills and 13 read-only MCP tools. “Read-only” describes this tool surface. An unrestricted agent with a shell or editor can still modify files through unrelated tools.

When the transport closes

The snapshot lives in the MCP process memory. If that transport dies, the old server cannot respawn itself inside the client. Reload or open a new task, call project_status, and index the selected project again.

Reference corpora are a separate lane

Operator-owned PDF material is built and attached separately. The corpus preserves page citations but does not download or redistribute documents for the user. A Graphify overlay is optional discovery evidence; its inference does not become a normative citation.

A responsible setup never promises “paste the link and it works.” It states the prerequisites, verifies the live MCP connection, and asks which MQL5 project is authorized for analysis.