K Dense Byok Ai Scientist

---
name: k-dense-byok-ai-scientist
description: AI co-scientist desktop app with Claude Scientific Skills, multi-model support, and 250+ scientific databases via BYOK API keys
triggers:
  - set up K-Dense BYOK
  - configure Kady AI assistant
  - add scientific skills to my project
  - integrate K-Dense with my research workflow
  - use K-Dense BYOK with my API keys
  - run AI co-scientist locally
  - connect OpenRouter to K-Dense
  - add web search to Kady agent
---

# K-Dense BYOK AI Co-Scientist

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

K-Dense BYOK is an open-source desktop AI research assistant ("Kady") that runs locally, uses your own API keys (BYOK), and gives you access to 170+ scientific skills, 250+ scientific databases, and 500k+ Python packages. It supports 40+ AI models via OpenRouter and optionally delegates heavy compute to Modal cloud infrastructure.

## Architecture

Three services run together on your machine:

| Service | Port | Role |
|---------|------|------|
| Frontend (Next.js) | 3000 | Chat UI, file browser, file preview |
| Backend (FastAPI) | 8000 | Kady agent, expert coordination, tools |
| LiteLLM proxy | 4000 | Routes AI requests to OpenRouter models |

Expert/coding tasks always use Gemini CLI via OpenRouter regardless of the model selected in the UI dropdown.

## Prerequisites

- macOS, Linux, or Windows (via WSL)
- Git
- An OpenRouter API key — [openrouter.ai](https://openrouter.ai/)
- *(Optional)* Parallel API key for web search — [parallel.ai](https://parallel.ai/)
- *(Optional)* Modal credentials for remote compute — [modal.com](https://modal.com/)

## Installation

```bash
git clone https://github.com/K-Dense-AI/k-dense-byok.git
cd k-dense-byok

Configuration

Create kady_agent/.env:

# Required
OPENROUTER_API_KEY=$OPENROUTER_API_KEY

# Optional — enables web search
PARALLEL_API_KEY=$PARALLEL_API_KEY

# Internal service wiring — leave as-is
GOOGLE_GEMINI_BASE_URL=http://localhost:4000
GEMINI_API_KEY=sk-litellm-local

# Default model for Kady (changeable in UI dropdown)
DEFAULT_AGENT_MODEL=openrouter/google/gemini-3.1-pro-preview

# Optional — Modal remote compute
MODAL_TOKEN_ID=$MODAL_TOKEN_ID
MODAL_TOKEN_SECRET=$MODAL_TOKEN_SECRET

Starting the App

chmod +x start.sh
./start.sh

On first run, start.sh automatically installs:

• Python packages and dependencies
• Node.js packages
• Gemini CLI
• Scientific skills

App opens at http://localhost:3000. Stop with Ctrl+C.

Project Structure

k-dense-byok/
├── start.sh              # One-command launcher
├── server.py             # FastAPI backend
├── kady_agent/
│   ├── .env              # API keys
│   ├── agent.py          # Main Kady agent definition
│   └── tools/            # Web search, delegation, file ops
├── web/                  # Next.js frontend
└── sandbox/              # Local file workspace (auto-created)

Key Files to Understand