MyAIGist MCP Server

MCP server providing document intelligence and knowledge management for Claude Desktop and other MCP-compatible clients. Process documents, answer questions with RAG, and maintain a persistent knowledge base - all running locally.

Overview

MyAIGist MCP provides 10 powerful tools for document intelligence and knowledge management:

• Document Processing: PDF, DOCX, TXT, URLs, file attachments, and batch processing

• Q&A System: RAG-powered question answering across multiple documents

• Knowledge Management: Persistent vector storage with document tracking

Features

✅ MCP-compatible - Works with Claude Desktop, Cursor, and other MCP clients ✅ Claude-powered - Uses Claude Sonnet 4.5 for summarization and Q&A ✅ Local execution - Runs on your machine with no external infrastructure ✅ Persistent storage - Single vector store survives across sessions ✅ Multi-document RAG - Unlimited documents in knowledge base ✅ Simple setup - Install dependencies, configure API keys, and go

Installation

Prerequisites

• Python 3.8+

• Anthropic API key (for Claude text generation)

• OpenAI API key (for embeddings)

• MCP-compatible client (Claude Desktop, Cursor, etc.)

Setup

1. Install dependencies:

   cd /path/to/myaigist_mcp pip install -r requirements.txt

2. Configure environment:

   cp .env.example .env # Edit .env and add both API keys: # - ANTHROPIC_API_KEY (for Claude) # - OPENAI_API_KEY (for embeddings)

3. Configure your MCP client:

   Claude Desktop - Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

   { "mcpServers": { "myaigist": { "command": "python3", "args": ["/path/to/myaigist_mcp/server.py"] } } }

   Cursor - Add to your Cursor MCP settings:

   { "mcpServers": { "myaigist": { "command": "python3", "args": ["/path/to/myaigist_mcp/server.py"] } } }

   Other MCP clients - Refer to your client's documentation for MCP server configuration.

4. Restart your MCP client

   The MCP server will start automatically when you open your client.

Architecture

Architecture Notes:

• All agents are self-contained in mcp_agents/

• No external dependencies on other projects

• Single-user design (no session/user isolation)

• Persistent vector storage with unlimited documents

• Claude Sonnet 4.5 for text generation (summarization, Q&A)

• OpenAI for embeddings (RAG)

Available Tools (10 Total)

Content Processing (5 tools)

1. process_document

Process a document from LOCAL FILE PATH (PDF, DOCX, TXT) and add to knowledge base.

Parameters:

• file_path (string, required): LOCAL filesystem path (e.g., /Users/mike/file.pdf)

• title (string, optional): Document title (defaults to filename)

• summary_level (string, optional): quick, standard, or detailed

Example:

Compatibility: Works with all MCP clients.

2. process_uploaded_document

Process a document attached to Claude Desktop (optimized for Claude Desktop file uploads).

Parameters:

• content (string, required): Text content extracted by Claude

• filename (string, required): Original filename

• summary_level (string, optional): quick, standard, or detailed

Example:

Compatibility: Designed for Claude Desktop. Other clients should use process_document with file paths.

3. process_text

Process raw text and add to knowledge base.

Example:

Compatibility: Works with all MCP clients.

4. process_url

Crawl web URL, extract content, and add to knowledge base.

Example:

Compatibility: Works with all MCP clients.

5. process_batch

Process multiple files and generate unified summary.

Example:

Compatibility: Works with all MCP clients.

Q&A System (1 tool)

6. ask_question

Ask questions about stored documents using RAG.

Example:

Compatibility: Works with all MCP clients.

Document Management (3 tools)

7. list_documents

List all documents in knowledge base with metadata.

Example:

Compatibility: Works with all MCP clients.

8. delete_document

Delete specific document by ID.

Example:

Compatibility: Works with all MCP clients.

9. clear_all_documents

Clear entire knowledge base.

Example:

Compatibility: Works with all MCP clients.

Utility Tools (1 tool)

10. get_status

Get system status and knowledge base statistics.

Example:

Compatibility: Works with all MCP clients.

Common Workflows

File Upload (Claude Desktop)

Single Document Q&A (Any MCP Client)

Multi-Document Research (Any MCP Client)

Configuration

Environment Variables (.env)

Model Selection

Current Configuration:

• Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) - Summarization and Q&A

• text-embedding-3-large - Higher accuracy for RAG vector search

Why Two API Keys?

• Anthropic Claude: Text generation (summarization, Q&A) - excellent quality

• OpenAI: Embeddings (Claude doesn't provide embeddings)

Storage

Vector Store:

• Path: data/vector_store.pkl

• Format: Pickle with numpy arrays

• Persistence: Survives server restarts

• Capacity: Unlimited documents

API Costs

Usage-based pricing:

• Claude Sonnet 4.5: $3/$15 per 1M tokens (input/output)

• OpenAI Embeddings: $0.13 per 1M tokens

Typical monthly usage:

• 100 documents processed: ~$5-10

• 500 questions answered: ~$2-5

• Total: ~$10-15/month

Troubleshooting

Server won't start

Import errors

API Key errors

Empty knowledge base after restart

• Check data/vector_store.pkl exists

• Verify file permissions (readable/writable)

• Check for errors in server logs

File upload not working

• Claude Desktop: Use process_uploaded_document for file attachments

• Other clients: Use process_document with local filesystem paths

• See tool descriptions for compatibility details

Development

Running Tests

Debugging

Project Structure

MCP Client Compatibility

Tested Clients

• ✅ Claude Desktop - Full support including file attachments

• ⚠️ Cursor - Core functionality works (use file paths instead of attachments)

• ❓ Other MCP clients - Should work with file path-based tools

Compatibility Notes

• All tools work with standard file paths

• process_uploaded_document is optimized for Claude Desktop's file attachment behavior

• Other clients should use process_document with local file paths

• Standard MCP protocol (stdio transport, JSON-RPC)

License

MIT License

Support

For issues or questions:

1. Check troubleshooting section above

2. Review MCP server logs (location varies by client)

3. Verify environment variables in .env

4. Test agent imports individually

Last Updated: 2026-01-19 Project Status: ✅ Production ready - 10 core tools implemented and tested Models: Claude Sonnet 4.5 (text) + OpenAI (embeddings) Focus: Document intelligence (text-based content only) Compatibility: Claude Desktop, Cursor, and other MCP-compatible clients