node version

# OCR Provenance MCP

**Your AI can't read documents. This fixes that. One command. Zero configuration.** OCR Provenance MCP gives your AI complete document intelligence — 153 tools for OCR, semantic search, vision AI, compliance, and provenance tracking — all running on local GPU. No cloud APIs. No data leaves your machine. Ever. ```bash npx -y ocr-provenance-mcp install # Docker (Linux/WSL + NVIDIA GPU) npx -y ocr-provenance-mcp install --mac # macOS (Apple Silicon, no Docker) ``` Works with Claude Code, Claude Desktop, Cursor, and Windsurf. No GPU? [Deploy on a cloud GPU in 4 steps.](#cloud-deploy-no-local-gpu-required) > **After installing, tell your AI: "Open the OCR Provenance dashboard."** The dashboard is your command center — view documents, manage databases, add funds, and monitor processing in real time at `http://localhost:3367`. ### NEW: Legal-Domain Embedding Model General-purpose AI search struggles with legal language. "Consideration" in a contract means contractual value, not thought. Case citations, statutory references, and legal terminology need a model trained on that vocabulary. OCR Provenance now ships with a **legal-domain embedding model** ([FreeLawProject ModernBERT](https://huggingface.co/freelawproject/modernbert-embed-base_finetune_8192)) fine-tuned on court opinions from CourtListener and RECAP. Choose it when creating a database for legal documents: ``` ocr_db_create { name: "vendor-contracts-2025", embedding_model: "legal" } ``` After that, all search, RAG, and embedding operations on that database automatically use the legal model. No extra configuration. [See details below.](#legal-document-intelligence) --- ## Install ### Docker (Linux / WSL2 + NVIDIA GPU) Requirements: [Docker Desktop](https://www.docker.com/products/docker-desktop/), [Node.js](https://nodejs.org/) 20+, NVIDIA GPU recommended (CPU works, just slower). ```bash npx -y ocr-provenance-mcp install ``` That's it. The installer: 1. Pulls the AI models image (~14 GB, cached after first install) 2. Pulls the application image (~500 MB) 3. Starts the container with automatic GPU detection 4. Provisions your license key 5. Registers with any detected AI client (Claude Code, Claude Desktop, Cursor, Windsurf) Restart your AI client. Then tell your AI: **"Open the OCR Provenance dashboard"** — this opens the web UI where you can view documents, manage databases, add funds, and monitor everything. ### macOS (Apple Silicon — No Docker Required) **Requirements:** - [Node.js](https://nodejs.org/) 20+ (`brew install node`) - Python 3.11+ (`brew install python@3.12`) - Apple Silicon Mac (M1/M2/M3/M4) with 12GB+ unified memory - ~15 GB free disk space (for AI models, one-time download) ```bash npx -y ocr-provenance-mcp install --mac ``` The installer: 1. Creates a Python virtual environment at `~/.ocr-provenance-mcp/venv/` 2. Installs PyTorch with MPS (Metal Performance Shaders) GPU acceleration 3. Installs Marker OCR, Chandra OCR 2, nomic embeddings, and all dependencies 4. Downloads model weights (~14 GB total, one-time, cached at `~/.cache/huggingface/hub/`) 5. Installs the dashboard web UI for account management and billing 6. Registers with any detected AI client (Claude Code, Claude Desktop, Cursor, Windsurf) Restart your AI client. The first OCR run downloads the Marker/Surya models (~3.3 GB, one-time). **What runs on Mac:** When your AI client connects, three services start automatically: - **MCP Server** — 153 tools via stdio (direct connection to your AI client) - **License Server** — billing, auth, Stripe payments (port 3000) - **Dashboard** — web UI at `http://localhost:3367` for account management, document viewing, and adding funds All three start together when your AI connects and stop together when it disconnects. **Optional:** Install [LibreOffice](https://www.libreoffice.org/) for Office-to-PDF conversion in the document viewer: ```bash brew install --cask libreoffice ``` **GPU memory guide:** | Mac | Unified Memory | What Works | |-----|---------------|------------| | M1/M2 (8 GB) | 8 GB | OCR + embeddings + search (VLM skipped automatically) | | M1/M2 Pro (16 GB) | 16 GB | Full pipeline. VLM may be skipped on large documents — close other apps for best results | | M3/M4 Pro (18-24 GB) | 18-24 GB | Full pipeline, comfortable headroom | | M4 Max/Ultra (48-192 GB) | 48-192 GB | Full pipeline, fast processing | **Where data is stored on Mac:** ``` ~/.ocr-provenance-mcp/ ├── venv/ # Python virtual environment ├── databases/ # SQLite databases (your documents, chunks, embeddings) ├── dashboard/ # Dashboard web UI └── wrapper.json # Configuration (license key, settings) ~/.cache/huggingface/hub/ ├── models--datalab-to--chandra-ocr-2/ # VLM model (~10 GB) └── models--nomic-ai--nomic-embed-text-v1.5/ # Embedding model (~500 MB) ``` **Mac troubleshooting:** | Issue | Fix | |-------|-----| | `Python 3.11+ required` | `brew install python@3.12` | | `No module named torch` | Re-run `npx -y ocr-provenance-mcp install --mac` | | VLM skipped (8 GB Mac) | Normal — OCR and search still work, just no image descriptions | | First OCR run is slow | Marker models downloading (~3.3 GB), subsequent runs are fast | | Dashboard not opening | Visit `http://localhost:3367` manually in your browser | ### Linux Bare-Metal (No Docker) Same as macOS but for Linux machines without Docker. Auto-detects NVIDIA CUDA GPUs and falls back to CPU. **Requirements:** [Node.js](https://nodejs.org/) 20+, Python 3.11+, ~15 GB free disk space. ```bash npx -y ocr-provenance-mcp install --bare ``` ### Cloud Deploy (No Local GPU Required) No GPU? No problem. Deploy on [RunPod](https://runpod.io) with a cloud GPU in 4 steps. #### Step 1: Create a Template Go to [RunPod](https://runpod.io) → **My Templates** → **New Template**. Fill in: - **Compute type:** Nvidia GPU - **Container image:** `ocrprovenance/ocr-provenance-mcp:latest` - **Container disk:** 20 GB - **Volume disk:** 100 GB - **Volume mount path:** `/data` - **HTTP ports:** `3000` (License Server), `3366` (MCP Server API), `3367` (Dashboard) - **Environment variables:** `MCP_TRANSPORT=http`, `TORCH_DEVICE=auto`, `MCP_HTTP_PORT=3366`, `PORT=3367` Click **Save Template**.

RunPod Create Template
Template settings with container image, ports, and environment variables

#### Step 2: Deploy a Pod Go to **Pods** → **Deploy**. Select your template and pick a GPU: | GPU | VRAM | On-Demand | Spot | Best For | |-----|------|-----------|------|----------| | RTX 3090 | 24 GB | ~$0.34/hr | ~$0.20/hr | Budget processing | | RTX 4090 | 24 GB | ~$0.44/hr | ~$0.28/hr | Fast processing | | RTX 5090 | 32 GB | ~$0.89/hr | ~$0.53/hr | Large docs + full VLM | Set **GPU count** to 1. Click **Deploy**.

RunPod Configure Deployment
GPU selection, pricing, and storage configuration

#### Step 3: Access Your Dashboard Wait ~2-3 minutes for startup. Click the **3367** port link to open the dashboard. All 153 MCP tools are available via the **3366** port.

RunPod Template Settings
Final template with ports and environment variables configured

Your data persists on the volume across pod stop/start. **Stop** (don't delete) the pod when not using it — you only pay while the GPU is running.

AI-Assisted Setup — Give this prompt to any AI agent with browser access to set it up for you

``` Go to https://runpod.io and set up a GPU pod: 1. Go to My Templates → New Template with these settings: - Template Name: OCR Provenance MCP - Compute type: Nvidia GPU - Container Image: ocrprovenance/ocr-provenance-mcp:latest - Container Disk: 20 GB - Volume Disk: 100 GB - Volume Mount Path: /data - HTTP Ports: 3000 (License Server), 3366 (MCP Server API), 3367 (Dashboard) - Environment Variables: MCP_TRANSPORT=http, TORCH_DEVICE=auto, MCP_HTTP_PORT=3366, PORT=3367 - Save the template 2. Go to Pods → Deploy: - Select the "OCR Provenance MCP" template - Choose RTX 5090 (32 GB) GPU, GPU count 1, On-Demand - Click Deploy 3. Wait for "Running" status, then click the 3367 port link to verify the dashboard loads. Click the 3366 port link and append /health to verify {"status":"ok"}. ```

Budget alternative: Vast.ai — RTX 3090 from ~$0.16/hr