Nutrient’s self-hosted MCP server lets agents process PDFs without sending them to the cloud

If your documents cannot leave your network, most agent demos stop at the upload button. Nutrient is positioning its Document Engine MCP Server for exactly that constraint: an MCP server that exposes document processing through natural language, either for a person in Claude Desktop or for an AI agent calling tools, while the engine itself runs on your infrastructure. The README draws a clear line between two Nutrient options. The DWS MCP Server is the cloud path, using a Nutrient-hosted API and an API key for workflows such as convert, OCR, redact, sign, extract, watermark, and automation. The Document Engine MCP Server is the self-hosted path, aimed at on-prem or private cloud deployments where deployment control and data residency matter more than avoiding operations. That distinction is the main decision to make before you look at setup. Running it locally first The documented setup is deliberately small. You clone the repo, start Nutrient Document Engine with Docker Compose, then point Claude Desktop at the MCP server through npx. The prerequisites listed are Claude Desktop, Docker Compose, and Node.js 22.x, 24.x, or 26.x at specific patched versions, because the server runs via npx as @nutrient-sdk/document-engine-mcp-server. The Claude Desktop config passes in the dashboard credentials, the engine base URL, and an API auth token. By default the engine runs with an evaluation license, and the README points to a separate section for removing evaluation watermarks and limitations. Documents are not uploaded through the chat itself in this flow. You open the dashboard at localhost:5100, drag and drop files in, then return to Claude Desktop to ask for things like listing documents and extracting text from a contract. The example conversations show the intended interaction pattern more than a fixed API surface.