Stache AI: Curated Storage for AI with Claude

Interested in how to give Claude access to some of your files?

I’ve been building Stache AI - curated storage for AI that uses MCP (Model Context Protocol) to connect directly with Claude.

The Problem

I got frustrated with AI assistants lacking access to my personal knowledge base. I wanted a solution that was simple to set up, worked directly with Claude, handled real documents, and didn’t require me to become an infrastructure expert.

What is Stache?

Stache is a locally-run document storage system designed to integrate with Claude via MCP. Instead of copying and pasting content into chat windows or dealing with context limits, you can ingest documents into Stache and let Claude search through them semantically.

Core features:

• Semantic Search - Find documents by meaning, not just keywords
• Local Deployment - Runs on your machine using Docker. Your data stays private.
• Multiple AI Backends - Supports OpenAI, Anthropic, Ollama, and AWS Bedrock
• Document Formats - PDF (with OCR), EPUB, Markdown, DOCX, PPTX, and transcript files (VTT/SRT)

Why MCP?

MCP (Model Context Protocol) is Anthropic’s open standard for connecting AI assistants to external tools and data sources. By building Stache with MCP support, Claude can:

• Search your knowledge base semantically
• Ingest text content directly
• List and retrieve documents
• Access your files without manual copy-paste

It works with both Claude Desktop and Claude Code.

Quick Setup

1. Start the server using Docker (with Ollama for local processing or OpenAI for cloud)
2. Install the MCP tools via pip or download the Windows executable
3. Add the server config to Claude Desktop or Claude Code

That’s it. No Kubernetes, no complex infrastructure.

Getting Started

Check out the Stache AI GitHub and the setup discussion for detailed instructions.

If you’re working with large document collections, technical documentation, or just want a better way to give Claude access to your knowledge - Stache might be what you’re looking for.

The project is MIT licensed.