Quick Links

Start	Setup	Tools	Security
Quick start	MCP client setup	Tool catalog	Security

What It Does

• Connects AI tools to a running local Power BI Desktop engine.
• Automates tables, columns, measures, relationships, DAX, refreshes, and Power Query.
• Reads/writes Excel files for local BI workflows.
• Extracts, patches, validates, and compiles report layouts through pbi-tools.

No Power BI Pro license is required for the local Desktop workflow.

Architecture

MCP Client --(stdio or sse)--> src/server.py
                              |
                              +-- src/mcp_core.py     FastMCP instance, _run, lifecycle
                              +-- src/wrappers/       14 thin wrappers/<domain>.py
                              +-- src/tools/          business logic (the *_tool fns)
                              +-- TOM/.NET ─────────> Power BI Desktop local SSAS
                              +-- ADOMD ────────────> DAX query execution
                              +-- openpyxl ─────────> Excel read/write/format
                              +-- pbi-tools ────────> report extract/compile + visuals
                              +-- src/security.py     path, DAX, payload safeguards

Full module layering and the visuals/ submodule fan-out: see ARCHITECTURE.md.

Requirements

Requirement	Install
Windows	Power BI Desktop local engine is Windows-only
Power BI Desktop	winget install Microsoft.PowerBIDesktop
Python 3.11+	winget install Python.Python.3.11
.NET 6+ Runtime	winget install Microsoft.DotNet.Runtime.6
pbi-tools	winget install pbi-tools or dotnet tool install -g pbi-tools

ADOMD.NET ships with Power BI Desktop. If pbi-tools is not on PATH, set PBI_TOOLS_PATH.

Quick Start

git clone https://github.com/imnotStealthy/powerbi-mcp-local.git
cd powerbi-mcp-local
pip install -r requirements.txt

Open Power BI Desktop with a .pbix file, then verify connectivity:

python tests/test_connection.py

Start the MCP server:

python src/server.py

Useful launch modes:

python src/server.py --transport sse --port 8765
python src/server.py --readonly
python src/server.py --profile readonly   # ~56 read tools
python src/server.py --profile write      # readonly + writes (no destructive)
python src/server.py --profile grading    # 25-tool surface for evaluation workflows
python src/server.py --profile all        # default — every tool

For SSE auth:

$env:PBI_MCP_AUTH_TOKEN = "your-secret-token"
python src/server.py --transport sse --port 8765

Clients must send:

Authorization: Bearer your-secret-token

MCP Client Setup

Standard stdio config:

{
  "mcpServers": {
    "powerbi": {
      "command": "python",
      "args": ["C:\\path\\to\\powerbi-mcp-local\\src\\server.py"]
    }
  }
}

Google Antigravity (Gemini-based IDE) ships a stricter MCP client that drops the connection at resources/list against the default FastMCP 1.27.x stdio server. Since v0.12.1 the project ships a dedicated entry point — src/server_antigravity.py — that strips the capability advertisement to tools only, forces UTF-8 / \n stdio, and routes every log to stderr at ERROR level. The default src/server.py is left untouched for Claude Desktop / Cursor / Anthropic CLI.

The PowerShell wrapper at tools/antigravity_mcp_launcher.ps1 pins the working directory and exports PYTHONUTF8=1 / PYTHONIOENCODING=utf-8 before invoking the venv Python on server_antigravity.py, so the encoding stays stable regardless of how Antigravity spawns the process.

%USERPROFILE%\.gemini\antigravity\mcp_config.json:

{
  "mcpServers": {
    "powerbi": {
      "command": "powershell.exe",
      "args": [
        "-NoProfile",
        "-ExecutionPolicy",
        "Bypass",
        "-File",
        "D:\\Users\\stealthy\\Documents\\GitHub\\powerbi-mcp-local\\tools\\antigravity_mcp_launcher.ps1",
        "--profile",
        "readonly"
      ]
    }
  }
}

Adjust the -File path to your clone. --profile readonly keeps tools/list small while validating the client connection — switch to --profile all (or omit the flag) once Antigravity lists tools successfully. --readonly and --profile {readonly,write,all,grading} are all forwarded through the wrapper to the entry point with the same semantics as src/server.py.

Local compatibility probe:

$init = @(
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"antigravity-probe","version":"1.0"}}}',
  '{"jsonrpc":"2.0","method":"notifications/initialized","params":{}}'
)

($init + '{"jsonrpc":"2.0","id":2,"method":"resources/list","params":{}}') -join "`n" |
  powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\tools\antigravity_mcp_launcher.ps1 --profile readonly

($init + '{"jsonrpc":"2.0","id":3,"method":"tools/list","params":{}}') -join "`n" |
  powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\tools\antigravity_mcp_launcher.ps1 --profile readonly

SSE endpoint:

http://localhost:8765/sse

Setup guides:

• docs/SETUP.md
• docs/WINDOWS_SETUP.md

Example Prompts

• Connect to Power BI and list all tables with columns.
• Create a measure called Total Sales in table Sales.
• Run this DAX query and show top 20 rows.
• Extract report, add a new page, place 3 visuals, then compile.

Tool Catalog

141 MCP tools are grouped into these areas. Visual creation is consolidated behind pbi_add_visual (28 chart types via visual_type) and pbi_add_visual_from_intent:

Area	Coverage
Model discovery	instances, tables, measures, relationships, metadata, validation
Model mutations	measures, columns, tables, relationships, formats, role-based DAX
Query and import	DAX execution, traces, validation, refresh, model export
Power Query (M)	read, write, create, import, bulk Excel/folder sources
PBIP/TMDL	list, read, write, and patch TMDL project files
Workflows	model audit, Excel import, measure workflow automation
Quality gates	DAX linting, visual checks, persistence, scenarios, report validation
RLS and calc groups	roles, filters, members, calculation groups
Excel	workbook, sheet, cell/range, formatting, search, Power BI import checks
Reports and visuals	extract, compile, pages, cards, charts, slicers, themes, dashboards

Unified visual creation is available through:

pbi_add_visual(visual_type, config)

In-place binding edits (no remove + recreate) via:

pbi_update_visual_bindings(extract_folder, page, visual_id,
                           projections={"Y": ["Total Sales"]})

Force Power BI Desktop to flush in-memory TOM mutations to the PBIX file (opt-in UI automation):

$env:PBI_MCP_ALLOW_UI_AUTOMATION = "1"   # set before launching the server

pbi_persist_now(pbix_path="C:\\reports\\sales.pbix", confirm=true)

Automation Flow

Excel input -> Power Query -> model updates -> measures -> validation -> report layout -> compile PBIX

Common tool chain:

excel_write_range
pbi_create_import_query
pbi_create_relationship
pbi_create_measure
pbi_refresh
pbi_execute_dax
pbi_extract_report
pbi_build_dashboard
pbi_compile_report

Troubleshooting

Symptom	Fix
No module named 'clr'	Install .NET 6+ runtime, then restart terminal
No running PBI Desktop instance found	Open a .pbix in Power BI Desktop first
pbi-tools not found	Add it to PATH or set PBI_TOOLS_PATH
PermissionError on .xlsx	Close Excel; workbook files are locked while open
Path blocked by policy	Configure PBI_MCP_ALLOWED_DIRS

Security

Built-in safeguards include:

• local path restrictions and traversal protection
• DAX/DMV unsafe-query guards
• Power Query SSRF protections (cloud DW + SaaS + reflection blocked by default)
• export redaction controls
• zip safety checks (Excel + PBIX)
• response-size cap (16 MiB by default) + per-minute rate limit (600/min)
• SSE Bearer auth + DNS-rebinding host/origin allowlist
• tool-call auditing

Details: SECURITY.md

Environment variables (quick reference)

Variable	Default	Purpose
PBI_MCP_AUTH_TOKEN	unset	SSE Bearer token (≥32 chars when set)
PBI_MCP_ALLOWED_ORIGINS	unset	Extra Host/Origin allowlist for SSE
PBI_MCP_ALLOW_UNAUTHENTICATED_SSE	0	Acknowledge non-loopback SSE without auth
PBI_MCP_READONLY	0	Block all write/destructive tools
PBI_MCP_ALLOW_DMV	0	Allow $SYSTEM.* / DISCOVER_* DAX
PBI_MCP_ALLOW_EXTERNAL_M	0	Bypass the M function blocklist
PBI_MCP_ALLOW_UI_AUTOMATION	0	Required for pbi_persist_now (Ctrl+S)
PBI_MCP_PERSIST_USE_SENDINPUT	0	Fall back from PostMessage to SendInput
PBI_MCP_PBI_TOOLS_TIMEOUT	300	pbi-tools subprocess timeout (s)
PBI_MCP_ALLOWED_DIRS	cwd	;-separated filesystem roots
PBI_MCP_SECURITY_POLICY	unset	Path or inline JSON for security_policy.json
PBI_MCP_AUDIT / PBI_MCP_STRICT_REGISTRY	0	Tool registry audit (CI)

Development

pip install -e ".[dev]"
pytest -q
ruff check src tests
ruff format --check src tests

CI runs pytest --cov=src --cov-fail-under=54 on Windows + an offline subset on Ubuntu, plus ruff lint + format check, on every PR. Strict registry audit (PBI_MCP_STRICT_REGISTRY=1) ensures every public pbi_*_tool has a matching @mcp.tool() wrapper.

Repository Layout

powerbi-mcp-local/
├── src/
│   ├── server.py            CLI + transport launcher + @mcp.resource/@mcp.prompt (~320 L)
│   ├── mcp_core.py          FastMCP instance + CONNECTION_MANAGER + lifecycle (~250 L)
│   ├── pbi_connection.py    TOM + ADOMD bring-up, write helpers, op history
│   ├── security.py          path / DAX / payload guards + tool category sets
│   ├── wrappers/            14 domain modules — `register_tool(pbi_*_tool)` calls
│   └── tools/               business logic (*_tool functions)
│       └── visuals/         17 focused submodules (layout, bindings, containers, charts, …)
├── tests/                   167 offline unit tests (live-only scripts gitignored)
├── .github/workflows/ci.yml pytest + coverage + ruff on Windows + Ubuntu, py3.11/3.12
├── docs/, specs/
├── ARCHITECTURE.md          module layering, visuals/ tree, profiles, registry audit
├── CHANGELOG.md             active changelog (last 3 releases)
├── CHANGELOG-archive.md     historical changelog
├── SECURITY.md
├── pyproject.toml           ruff, pytest, coverage config + dev deps
└── requirements.txt

License

MIT