Summary

K-Dense BYOK provides a desktop AI co-scientist workspace for researchers and analysts, solving vendor lock-in by enabling users to integrate their own API keys from diverse AI providers. It leverages specialized AI "experts" and an extensive library of scientific skills, offering a powerful, customizable AI workspace without provider constraints.

How It Works

The application comprises three local services: a web frontend, a backend orchestrating the Kady agent and expert tasks, and a LiteLLM proxy for AI model routing. Kady interprets prompts, delegating complex work to specialized AI experts equipped with over 170 scientific skills, 250+ databases, and 500k+ Python packages. The "Bring Your Own Keys" (BYOK) model emphasizes user control over AI model selection and costs via OpenRouter, integrating web search and local file handling within a secure sandbox.

Quick Start & Requirements

Installation involves cloning the GitHub repository and running start.sh. Requirements include macOS/Linux (WSL for Windows) and an OpenRouter API key. Optional keys for Parallel.ai (web search) and Modal (remote compute) enhance functionality. The initial setup automatically installs dependencies like Python packages, Node.js, and Gemini CLI, taking a few minutes. Links to OpenRouter, Parallel, and Modal are provided.

Highlighted Details

• Supports 40+ AI models from major providers (OpenAI, Anthropic, Google, etc.) via OpenRouter.
• Integrates 170+ scientific skills, 250+ scientific databases, and 500k+ Python packages.
• Features a local sandbox for secure file operations across nearly all file types.
• Offers optional remote computation via Modal for demanding workloads.
• Employs expert delegation for specialized AI tasks (bioinformatics, data analysis).

Maintenance & Community

Currently in beta with ongoing development. Community engagement is via GitHub issues for bug reports/feature requests. Updates and tutorials are shared on X, LinkedIn, and YouTube.

Licensing & Compatibility

The specific open-source license is not explicitly stated. Compatibility focuses on local desktop execution on macOS, Linux, and Windows (via WSL). The BYOK approach ensures broad compatibility with AI model providers.

Limitations & Caveats

The project is in beta with planned improvements. Expert execution and coding tasks consistently use Gemini models via Gemini CLI, regardless of the main agent's selection. Windows users require WSL. The absence of a stated license requires further investigation for commercial use.