Automated arXiv research paper discovery and AI-powered summarization system with MCP (Model Context Protocol) integration for Claude Desktop. Built as a vibe-coding project, we keep the tooling lightweight and the feedback loop fast so you can stay in flow while shipping research insights.

Features

Daily Automation: Automatically searches arXiv for papers matching your topics
AI Summaries: Claude-powered intelligent paper summarization with cost optimization
Multi-User: Support for multiple users with independent topic configurations
MCP Integration: Direct integration with Claude Desktop via Model Context Protocol
RSS Feeds: Per-user RSS/Atom feeds for any feed reader
Web Dashboard: Optional SvelteKit dashboard (coming soon)
Cost-Effective: Runs mostly on Cloudflare free tier (~$3/month for 2 users)

Architecture

MCP Server: TypeScript Worker handling MCP protocol and tool execution
Daily Automation: Cron-triggered Worker for paper collection and summarization
Storage: Cloudflare D1 (SQLite), R2 (PDFs), KV (cache)
AI: Claude 3.5 Sonnet with Haiku triage for cost optimization

Quick Start

Prerequisites

Cloudflare account with Workers, D1, KV, and R2 enabled
Anthropic Claude API key (console.anthropic.com)
Bun 1.1+ (package manager - install from bun.sh)
Wrangler CLI (installed automatically via bun)

Installation

# Clone the repository
git clone https://github.com/jeffaf/kivv.git
cd kivv

# Install dependencies with bun (our vibe-coding default)
bun install

# Set up environment variables
cp .env.example .env
# Edit .env with your API keys and Cloudflare credentials

Deployment

For complete deployment instructions, see DEPLOYMENT.md.

Quick deployment:

# Set secrets for automation worker
cd automation
wrangler secret put CLAUDE_API_KEY
wrangler secret put CRON_SECRET

# Deploy automation worker
wrangler deploy

# Deploy MCP server
cd ../mcp-server
wrangler deploy

# Configure Claude Desktop (see DEPLOYMENT.md for details)
# Add MCP server URL and API key to Claude Desktop config

Verification

# Test automation worker
curl https://kivv-automation.<username>.workers.dev/health

# Test MCP server
curl https://kivv-mcp.<username>.workers.dev/health

# Check database
wrangler d1 execute kivv-db --command "SELECT COUNT(*) FROM users"

# Or use the automated health check script
./scripts/health-check.sh

See TROUBLESHOOTING.md if you encounter any issues.

Automated Deployment Script

For one-command deployment:

# Run the automated deployment script
./scripts/deploy.sh

# This will:
# - Verify prerequisites
# - Check infrastructure
# - Configure secrets
# - Deploy both workers
# - Provide Claude Desktop config

Project Structure

kivv/
├── mcp-server/          # MCP Server Worker (Claude integration)
├── automation/          # Daily automation Worker (cron)
├── shared/              # Shared types and utilities
├── tests/
│   ├── security/       # Security tests (auth, injection, XSS)
│   ├── integration/    # MCP tool integration tests
│   └── unit/           # Unit tests
├── .checkpoint/         # Development checkpoints
└── package.json         # Monorepo root with workspaces

Documentation

Deployment Guide - Complete production deployment instructions
Troubleshooting Guide - Common issues and solutions
Setup Checklist - Infrastructure setup
Implementation Plan - Chunked development guide
API Documentation - API endpoints and usage
PRD (Full Spec) - Complete technical specification

CI & Automation

Tests & Type Checks: GitHub Actions run bun run type-check and bun test on pushes and pull requests to keep the vibe coding fast without breaking the build.
Deployments: Cloudflare Workers deploy pipelines trigger when mcp-server/ or automation/ change, using wrangler with your configured Cloudflare credentials.

Development

# Run type checking
bun run type-check

# Run all tests
bun test

# Run security tests specifically
bun run test:security

# Run tests in watch mode
bun run test:watch

# Run MCP server locally
bun run dev:mcp

# Run automation worker locally
bun run dev:automation

# Build all workspaces
bun run build

Testing

The project includes comprehensive test coverage:

Security Tests (100% coverage required):
- Authentication (API key validation)
- Authorization (user data isolation)
- SQL injection prevention
- XSS prevention in RSS feeds
- Rate limiting enforcement
Integration Tests: End-to-end MCP tool workflows
Unit Tests: Isolated utility function testing

# Run with coverage report
bun run test:coverage

License

MIT

Kivv