A desktop AI co-scientist for researchers with commitment issues

What it does

K-Dense BYOK is a free, open-source desktop app that runs locally on macOS or Linux and gives you an AI research assistant named Kady. You chat with Kady as you would any assistant, but for heavier work it spins up specialized expert agents behind the scenes, each equipped with a full Python environment and a library of over 140 pre-installed scientific skills. The app connects to tool-capable models through OpenRouter or runs local models via Ollama, and it can tap into 229 scientific and financial databases and 326 workflow templates spanning fields from genomics to astrophysics.

The interesting bit

The architecture treats your desktop like a shared lab bench: you can run up to ten chat tabs in parallel, and every tab shares the same project sandbox, so a file written by an expert in one tab is immediately visible in the others. When the work is done, a one-click Copy as Methods button spits out a publication-ready paragraph summarizing the session provenance—an unusually thoughtful nod to actual scientific workflow.

Key highlights

• Model-agnostic by design: pick any tool-capable model from OpenRouter (OpenAI, Anthropic, Google, xAI, Qwen, etc.) or run free local models through Ollama, with separate orchestrator and expert selections per chat tab.
• 140+ scientific skills pre-installed covering genomics, proteomics, drug discovery, and materials science, plus 229 databases across 18 categories including biomedical, chemistry, earth and climate, and stock market data.
• Rich file previews for scientific formats (FASTA, FASTQ, VCF, BED, GFF, SAM, BCF), Jupyter notebooks, CSVs, PDFs, and a split-pane LaTeX editor with live compilation.
• Optional remote compute via Modal for cloud GPU jobs (T4 through H100) without leaving the input bar.
• Extensible via MCP servers and browser automation; includes web search through Exa or Parallel and biomedical literature search through Paperclip.

Caveats

• The project is explicitly in beta, and the README points to a dedicated limitations document with rough edges, especially around the expert system.
• Windows support is only through WSL, not native.
• Many features are still being refined, including better skill utilization and a faster PDF parser.

Verdict

Worth a look if you are a scientist, analyst, or bioinformatician who wants a local, extensible AI workspace without surrendering your data or your choice of model provider. Skip it if you need a polished, zero-setup consumer tool or native Windows support.