document-intelligence

# Document Intelligence — MCP Workspace AI-powered document analysis built on the [Model Context Protocol](https://modelcontextprotocol.io). Drop a PDF or image, run one or several analysis tools, get structured JSON — classification, entity extraction, summarization, validation, and comparison, all powered by Claude. <div align="center"> [](https://mcp-document-front.vercel.app/) </div> <div align="center"> <a href="https://mcp-document-front.vercel.app/"> <img src="./assets/demo.gif" alt="Document Intelligence demo" width="800"> </a> </div> --- ## Projects This workspace contains two independent Next.js applications: | Project | Description | Default URL | |---------|-------------|-------------| | [`mcp-document`](https://github.com/P4ST4S/mcp-document) | MCP server — exposes the 5 analysis tools via JSON-RPC | `http://localhost:3000` | | [`mcp-document-front`](https://github.com/P4ST4S/mcp-document-front) | Web UI — drag & drop interface, proxies requests to the MCP server | `http://localhost:3001` | --- ## How it works ``` User (browser) └── uploads a file → mcp-document-front (Next.js) └── POST /api/analyze → mcp-document (MCP server) └── Claude API (Haiku / Sonnet) ``` The frontend never calls the Claude API directly. Every request is proxied through the MCP server, with the user's API key forwarded via a secure `x-api-key` header (decrypted server-side from an `httpOnly` cookie). --- ## Authentication There is no shared API key. Each user enters their own [Anthropic API key](https://console.anthropic.com/settings/keys) on first visit. The key is: 1. Validated against the Anthropic API before being saved 2. Encrypted with AES-GCM 256-bit using a server-side `SESSION_SECRET` 3. Stored in an `httpOnly; SameSite=Strict; Secure` cookie (7-day session) 4. Never exposed to the browser — not in JS, not in localStorage, not in response headers The MCP server also accepts a fallback `ANTHROPIC_API_KEY` environment variable for Claude Desktop and direct MCP usage. --- ## Tools | Tool | Model | Description | |------|-------|-------------| | `classify_document` | Haiku | Categorize: invoice, contract, identity document, bank statement, quote, other | | `extract_entities` | Haiku | Named entities: dates, amounts, names, SIRET, IBAN, addresses, signatories | | `summarize_document` | Sonnet | Structured summary + key points (French or English) | | `validate_document` | Haiku | Completeness check against an expected document type | | `compare_documents` | Sonnet | Diff two documents — differences, similarities, recommendation | All tools accept PDF, PNG, JPEG, or plain text as base64-encoded input (max 4 MB). --- ## Quick Start ```bash # 1. Start the MCP server cd mcp-document pnpm install cp .env.local.example .env.local # ANTHROPIC_API_KEY is optional (Claude Desktop fallback) pnpm dev # → http://localhost:3000 # 2. Start the frontend (new terminal) cd mcp-document-front pnpm install cp .env.local.example .env.local # Edit .env.local: # SESSION_SECRET=$(openssl rand -hex 32) # MCP_SERVER_URL=http://localhost:3000/api/mcp pnpm dev # → http://localhost:3001 ``` Open `http://localhost:3001`, enter your Anthropic API key, and drop a document. --- ## Deployment Both projects deploy independently to Vercel. The frontend points to the deployed MCP server via `MCP_SERVER_URL`. See each project's README for full deployment instructions: - [mcp-document → Deploy to Vercel](https://github.com/P4ST4S/mcp-document/blob/main/README.md#deploy-to-vercel) - [mcp-document-front → Deploy to Vercel](https://github.com/P4ST4S/mcp-document-front/blob/main/README.md#deploy-to-vercel) --- ## Tech Stack - **Next.js 16** (App Router) — both projects - **@anthropic-ai/sdk** `^0.97.1` — Claude API client - **@modelcontextprotocol/sdk** + **mcp-handler** — MCP server infrastructure - **React 19**, **TypeScript 5**, **Tailwind CSS 4** - **Web Crypto API** — AES-GCM 256-bit encryption (no extra dependencies)