PBIFORGE MCP Server

Local Claude Desktop MCP app for generating Power BI .pbit reports from CSV datasets, natural-language instructions, and reference-image styling.

GitHub

Documentation

ReportForge PBI

Generate Power BI (.pbit) dashboards from CSV files or live SQL databases — locally, on your own machine, in seconds.

ReportForge profiles your data, designs a sensible dashboard layout (KPI cards + charts + filters), compiles it to a Power BI template, and writes the result to disk. Open the .pbit in Power BI Desktop and you have a working report.

Highlights

CSV upload — drop a single .csv, get a .pbit. No Power BI required to generate it.
Live SQL sources — SQL Server, PostgreSQL, Oracle. The generated .pbit queries the database on open; credentials are never embedded.
Plain-English prompts — describe what you want ("bar chart company by revenue, line of growth over year"). Powered by Claude or GPT — bring your own API key. Without a key, ReportForge still produces a sensible default dashboard from a CSV profile; the prompt is ignored.
Layout from a reference image — drop in a screenshot of a dashboard you like. ReportForge reads the structure (card count, chart types, filter panel) and applies it to the generated layout, then adopts the color palette as the report theme.
Multiple chart types — bar, column, line, area, pie, donut, scatter, table. KPI cards. Slicers.
Browser UI — runs on http://127.0.0.1:8000/. Local-only by default.
MCP server — exposes the same tools to Claude Desktop over stdio for "/" prompts inside Claude.

Quick start

# 1. Clone
git clone https://github.com/twilize5/reportforge.git
cd reportforge-pbi

# 2. Install Python deps (creates .venv)
.\setup_local.ps1

# 3. Install pbi-tools.core somewhere on PATH or at C:\pbi-tools\
#    Download: https://github.com/pbi-tools/pbi-tools/releases

# 4. Start the local server
.\run_api.ps1

Open http://127.0.0.1:8000/ in your browser.

If you'd rather use ReportForge inside Claude Desktop, see LOCAL_SETUP.md for the stdio MCP setup.

The browser UI

Field	What it does
Data source	Pick `CSV file` or one of the SQL connectors.
CSV file	Single CSV with a header row. Columns are profiled automatically (measures, dimensions, dates, geographic fields).
SQL fields	`Server`, `Database`, `Schema`, `Table` (or a free-form `SELECT` query). Username/password used only for profiling — never embedded in the `.pbit`.
Prompt	Plain-English description of what you want to see. Requires an LLM key (see "AI assistance" below). Without a key the prompt is ignored and you get the deterministic default dashboard.
Style reference image	Optional PNG/JPG. ReportForge reads the dashboard structure (number of KPI cards, chart types, filter panel presence) using Claude Vision and mirrors that layout in the generated report. Colors are also extracted and applied as the report theme.
AI assistance	Required for prompts to take effect. Pick Anthropic/OpenAI and paste an API key, or pick local Qwen through Ollama. Only column names (not data rows) are sent to the model.

The downloaded .pbit opens in Power BI Desktop, prompts for any data-source credentials, and renders the dashboard.

How prompts work

ReportForge always runs the deterministic dashboard builder (KPI cards, headline bar/line chart, donut for low-cardinality dims, second "Breakdowns" page when the dataset is rich). On top of that, if you supply an LLM key, your prompt is interpreted and additional visuals are appended.

What the LLM sees:

Your prompt
Column names + roles (measure / dimension)
Column semantic types (temporal, geographic, categorical, etc.)
Visuals the deterministic builder already produced (so it doesn't duplicate them)

What it does NOT see: any of your data rows.

The model returns a JSON list of chart specs (type, x-axis column, y-axis column, title); ReportForge validates each spec against the column profile and adds the ones that pass.

The response header X-ReportForge-Parser tells you which path ran for any given generate (llm:anthropic, llm:openai, llm:qwen, llm:<provider>-empty, no-key, or deterministic-only). The UI shows this under each "Report ready" message.

SQL data sources

ReportForge ships with lazy driver imports — the base install works for CSV-only use. Install drivers only for the databases you need:

# SQL Server (also needs the Microsoft ODBC Driver 17 or 18, installed system-wide)
.\.venv\Scripts\python.exe -m pip install pyodbc

# PostgreSQL
.\.venv\Scripts\python.exe -m pip install psycopg2-binary

# Oracle (thin mode - no Oracle client install required)
.\.venv\Scripts\python.exe -m pip install oracledb

In the UI, pick the source kind from the Data source dropdown. ReportForge:

Connects with the credentials you supply.
Pulls SELECT TOP 1000 (or LIMIT 1000) to profile column types, cardinality, and roles.
Generates a Power Query M expression that Power BI Desktop will use to fetch the full dataset live.
Compiles everything into a .pbit.

The credentials never make it into the .pbit file — Power BI Desktop will prompt for sign-in on first open. This is the standard Power BI Template behavior.

Example: SQL Server

Field	Example
Server	`db-prod.corp.com,1433` or `localhost\SQLEXPRESS`
Database	`AdventureWorks2022`
Schema	`Sales` (defaults to `dbo`)
Table	`Customer`
Username/Password	Leave blank for Windows auth, fill for SQL auth.

Example: free-form query

Paste a SELECT (with joins, filters, anything) into the Or: SELECT query box. ReportForge will profile the first 1000 rows of the result and target the same query in the generated M.

AI assistance

Provider	Model used	Where to get a key
Anthropic	Claude Sonnet with fallbacks	https://console.anthropic.com/
OpenAI	`gpt-4o-mini`	https://platform.openai.com/api-keys
Local Qwen	Auto-detected Ollama Qwen model, or `QWEN_MODEL` / `OLLAMA_MODEL`	No API key required

What gets sent to the model:

Your prompt
The list of column names + their inferred roles
The list of visuals already generated by the deterministic builder
Nothing else. Your data rows never leave your machine.

The key is cached in the browser tab's sessionStorage so you don't have to re-paste it on every generate. It's wiped when the tab closes.

For local Qwen, choose Local Qwen 2.5 (Ollama) in the provider dropdown. The Windows installer bundles a CPU-only Ollama runtime and will start it when ReportForge launches; qwen2.5:3b is pulled on first use if it is missing, so the first Qwen generation can take a few minutes and needs internet access. ReportForge calls http://127.0.0.1:11434 by default; override with QWEN_OLLAMA_URL or OLLAMA_HOST if your Ollama server runs elsewhere.

If you skip the key, ReportForge falls back to the deterministic dashboard builder only. Your prompt has no effect in that case — the dashboard is built purely from the column profile.

Style reference image

Drop any PNG/JPG (a screenshot of a dashboard you like, a brand mockup, a Tableau page). ReportForge runs two extractions in parallel:

Layout extraction (Claude Vision, requires Anthropic key) Reads the structural layout of the reference dashboard:

How many KPI/metric cards are visible → used as the KPI card count in the generated report
Which chart types are present (bar, column, line, area, donut, etc.) → influences chart type selection
Whether a right-side filter/slicer panel exists → adds or omits a slicer panel
Whether the KPI cards are in a left-column rail or a top-row grid → picks the matching layout template

Palette extraction (Claude Vision, or local Pillow fallback) Extracts the color scheme:

Primary, secondary, and accent colors
Page canvas background
Chart series palette (8-color rotation)

The layout extraction requires an ANTHROPIC_API_KEY. If no key is available, ReportForge still samples colors locally with Pillow and the layout falls back to the data-driven default.

API endpoints

All endpoints are local-only (127.0.0.1:8000).

Method + path	Purpose
`GET /health`	Liveness probe.
`GET /`	The browser UI.
`POST /generate-from-csv`	Multipart: `csv_file`, `prompt`, `image_file?`, `llm_provider?`, `llm_api_key?`. Returns the `.pbit` file.
`POST /generate-from-source`	Multipart: `kind`, `server`, `database`, `schema`, `table` or `query`, `username`, `password`, `prompt`, `image_file?`, `llm_provider?`, `llm_api_key?`. Returns the `.pbit`.
`POST /generate`	LLM-only path (requires `ANTHROPIC_API_KEY` env var). For Claude Desktop.
`/mcp/*`	The MCP server, exposed over HTTP. Used by Claude Desktop and other MCP-aware clients.

For Claude Desktop integration over stdio, point the launcher at run_mcp_stdio.ps1 — see LOCAL_SETUP.md.

Windows installer (single .exe for end users)

A double-clickable installer is available for users who don't want to run setup_local.ps1. It bundles Python, all dependencies, the FastAPI app, and pbi-tools.core into one .exe. See installer/README.md for the build process. End-user install flow:

Download ReportForge-PBI-Setup-1.0.3.exe (or the latest version).
Run it (admin required — installs into Program Files).
Start Menu → ReportForge PBI. The server starts and your browser opens automatically.
User data (generated reports, sources, sessions) lives at %LOCALAPPDATA%\ReportForge\ and survives reinstall/uninstall.

Power BI Desktop External Tool button

Optional. Adds a ReportForge PBI button to the External Tools ribbon in Power BI Desktop. See EXTERNAL_TOOL_SETUP.md.

Some tenants block External Tools registration via Fabric tenant policy. If the ribbon button doesn't appear after registering, the browser UI still works — just open http://127.0.0.1:8000/ directly.

Architecture (one screen)

                ┌──────────────┐
   CSV / SQL → ┤ data_profiler / data_sources ┤ → DatasetProfile
                └──────────────┘
                        │
                        ▼
                  build_intent_from_profile      (deterministic)
                        │
                        │  + prompt parser (regex OR LLM)
                        │  + image layout hint extractor  (Claude Vision)
                        │  + image palette extractor      (Claude Vision / Pillow)
                        ▼
                  ReportIntent (Pydantic)
                        │
                        ▼
                  build_bim_from_profile  →  semantic model (BIM JSON)
                  build_m_for_source      →  Power Query M
                  build_layout_from_intent →  report layout (visuals + theme)
                        │
                        ▼
                  pbi-tools.core compile  →  .pbit
                  inject_data_mashup       (DataMashup binary)
                  inject_report_layout     (Report/Layout)
                        │
                        ▼
                      .pbit file

Key files:

main.py — FastAPI app, endpoints, static UI mount.
orchestrator.py — pipelines (pipeline_from_csv, pipeline_from_source).
data_profiler.py — CSV profiler.
data_sources.py — DB profilers + M generators for MSSQL / Postgres / Oracle.
auto_intent.py — deterministic intent + layout builder, regex prompt parser.
llm_intent.py — LLM-driven prompt parser (Anthropic + OpenAI).
image_analyzer.py — palette extraction and layout hint extraction from reference images.
file_writer.py — writes the project tree, injects DataMashup + Report/Layout.
mcp_server.py — MCP tools for Claude Desktop.
static/index.html — browser UI.

Troubleshooting

.pbit won't open / "An error occurred while opening the file". Usually a stale generation in generated_reports/. Delete and regenerate. If it persists, check the server logs — pbi-tools.core errors usually show up there.

External Tools button doesn't appear in Power BI Desktop. Most often a Fabric tenant policy. See the troubleshooting section in EXTERNAL_TOOL_SETUP.md.

LLM call fails silently. When the LLM call errors or returns an empty visuals list, ReportForge falls back to the deterministic builder and surfaces llm:<provider>-empty in the X-ReportForge-Parser response header (visible in the UI status message). Double-check the key value, the provider dropdown, and that openai is installed (pip install openai) if you picked OpenAI.

SQL profiling fails with "pyodbc/psycopg2/oracledb is required". Install the relevant driver into the venv (see SQL data sources).

Generation succeeds but the dashboard looks plain blue, not like my reference image. The local Pillow palette extractor is more conservative than the Anthropic vision path. If you have an ANTHROPIC_API_KEY set, the vision path gives better color results. Or open the resulting .pbit and tweak the theme manually.

Reference image supplied but the layout doesn't match it. Layout extraction requires an ANTHROPIC_API_KEY — without one, only colors are sampled and the layout falls back to the data-driven default. Check that the key is set and that the reference image clearly shows the dashboard structure (card tiles, chart panels, filter sidebar).

Project state

This is the local-first incarnation of ReportForge. Railway deployment is not the supported path right now — everything is designed to run on your machine, talk to local CSVs and databases, and write .pbit files into generated_reports/.

The roadmap (see .claude/plans/) covers:

Phase 1: SQL connectors (✅ this release).
Phase 2: joined multi-table models with auto-relationship inference.
Phase 3: read the model out of an open Power BI Desktop session via a .NET helper.