bibiibjorn-Finvision/MCP-PowerBi_Finvision

MCP-PowerBi-Finvision Server (v2.4)

Analyze your Power BI Desktop model locally with an MCP server. Browse schema, inspect DAX and M, run SE/FE performance checks, export docs, and more — just by asking your AI client.

• Status: Production-ready with security hardening • OS: Windows 10/11 • Last updated: 2025-10-10

Who is this for?

• Power BI analysts who want instant, safe model introspection from Claude or ChatGPT
• Engineers who need fast diagnostics, exports, and best-practices checks

Features at a glance

• Auto-detect and connect to Power BI Desktop
• List tables/columns/measures, preview data, inspect relationships and M
• Validate DAX, scan VertiPaq stats, analyze cardinality/encoding
• Detect unused objects, analyze dependencies and RLS coverage
• Export compact schema, TMSL/TMDL, relationship graph, docs
• NEW v2.4: Input validation, rate limiting, enhanced error context

Security & Performance (v2.4)

• ✅ Input sanitization - Prevents DAX/M injection attacks
• ✅ Rate limiting - Protects Desktop from query overload (10 req/sec default)
• ✅ Path validation - Prevents directory traversal in exports
• ✅ Cache management - Bounded memory (100MB/1000 entries)
• ✅ Enhanced errors - Desktop version context & known issue detection
• ✅ Tool timeouts - Per-tool execution limits

Requirements

• Windows 10/11 (64‑bit)
• Power BI Desktop (latest recommended)
• .NET Framework 4.7.2+
• An MCP-capable client: Claude Desktop or ChatGPT Desktop (with MCP)

Quick start (Windows)

1. Place this folder somewhere stable, e.g. C:\Tools\pbixray-mcp-server.

2. Create a Python virtual environment and install requirements.

• Open Windows PowerShell in the project folder and run the steps from INSTALL.md (section "Set up Python (venv)").

3. Optional: install the .NET assemblies for advanced features:

cd lib/dotnet
./install.ps1

4. Open Power BI Desktop with a .pbix loaded and wait ~10 seconds.

5. Connect your AI client:

• Claude Desktop (recommended): run ./scripts/install_to_claude.ps1, then fully restart Claude.
• ChatGPT Desktop: use the manual Developer JSON steps in INSTALL.md (no script required).

6. In your AI client, try: "Detect Power BI Desktop instances" → "Connect to instance 0" → "List tables".

You're ready to explore your model.

Common things to ask

• "Search for measures containing CALCULATE"
• "Analyze this DAX and break down SE vs FE"
• "Export a compact schema" or "Generate documentation"
• "Show me rate limit stats" (NEW)
• "Check cache performance" (NEW)

Full analysis: fast vs normal

• full_analysis: summary, relationships, best practices, M scan, optional BPA

  • profile: fast | balanced | deep (default: balanced)
  • depth: light | standard | deep (default: standard)
  • include_bpa: true/false (ignored when profile=fast)

• propose_analysis: returns a small menu with fast vs normal options when a user asks to "analyze the model".

Install and manage

• Full install/update/uninstall steps: see INSTALL.md
  • Scripted install (Claude)
  • Manual install for Claude and ChatGPT (Developer JSON)
• Script catalog and recommendations: see docs/Scripts.md

Docs and links

• Quickstart (friendly): docs/PBIXRAY_Quickstart.md
• Scripts catalog: docs/Scripts.md
• .NET notes: lib/dotnet/README.md, lib/dotnet/dotnet_versions.md
• Contributing & Git workflow: CONTRIBUTING.md

If you use GitHub Pages, .github/workflows/docs.yml publishes docs/ on push to main.

Tips, diagnostics, and safety

• Results are paged (page_size + next_token) for large outputs
• Use get_recent_logs, summarize_logs, and get_server_info for quick diagnostics
• NEW: Use get_rate_limit_stats and get_cache_stats for performance monitoring
• To force a fresh DAX run, set bypass_cache: true on run_dax
• For DMV queries with $SYSTEM.*, wrap sources in TOPN(...) before SELECTCOLUMNS to materialize

Privacy & security

• Everything runs locally over stdio; no ports exposed
• Connections use Desktop's embedded SSAS over localhost
• Input validation prevents injection attacks
• Rate limiting protects Desktop from overload
• Logs live in logs/pbixray.log

Error envelope shape (for integrators)

• Not connected → { success: false, error_type: "not_connected", ... }
• Manager unavailable → { success: false, error_type: "manager_unavailable", required_manager: "<n>" }
• Unknown tool → { success: false, error_type: "unknown_tool", tool_name: "..." }
• Unexpected error → { success: false, error_type: "unexpected_error", tool_name: "..." }
• NEW: Rate limited → { success: false, error_type: "rate_limited", retry_after_seconds: X }
• NEW: Validation error → { success: false, error_type: "validation_error", field: "..." }

Successful responses include minimal connection metadata when available, like { port: "<desktop-port>" }.

Configuration

Edit config/default_config.json or create config/local_config.json for overrides.

New v2.4 settings:

• rate_limiting.enabled - Enable rate limiting (default: true)
• rate_limiting.profile - conservative | balanced | aggressive | development
• security.enable_input_validation - Validate inputs (default: true)
• security.strict_m_validation - Strict M expression validation (default: false)

Support

• Ensure a .pbix is open, then detect → connect → run tools
• Verify .NET assemblies: venv\\Scripts\\python.exe scripts\\verify_dotnet_assemblies.py
• See INSTALL.md for troubleshooting and uninstall

Changelog (v2.4)

Added

• Input validation to prevent injection attacks
• Rate limiting (token bucket, per-tool limits)
• Enhanced error handler with Desktop version detection
• Cache manager with size limits and eviction metrics
• Tool-specific timeout configuration
• .NET assembly auto-installer and verifier
• Monitoring tools: get_rate_limit_stats, get_cache_stats, get_error_stats

Security

• DAX/M expression validation
• Path traversal prevention
• Identifier sanitization
• Configurable rate limits

Performance

• Bounded cache (100MB, 1000 entries, LRU eviction)
• Per-tool timeouts
• Cache hit rate tracking

— Happy analyzing!