K-Dense has released K-Dense BYOK (Bring Your Own Keys), a free, open-source AI research assistant built for scientists across any field. The software runs locally on a user's computer and connects to AI models through accounts and API keys the user already holds. As of September 11, 2026, the platform holds the #2 position on the Genentech CompBioBench v1 public leaderboard, a benchmark of 100 computational-biology tasks, where it scored 98% accuracy using GPT-6 Astra with xhigh reasoning.
The assistant, called Kady, works through research tasks described in plain language. A scientist can ask it to analyze a dataset, review a manuscript, search the literature, or build a figure. Kady writes and runs analysis code, searches and reads sources, creates figures and reports, and maintains a record of its actions. The workspace stays entirely on the user's machine - K-Dense does not host or store project files. When a hosted AI model is used, material for that request is sent directly to the selected provider under that provider's privacy terms.
How the workspace operates
Kady is designed to carry out research work rather than just answer questions. It can clean data, run statistical methods, compare groups, fit models, and generate publication-ready figures. It searches the web and reads PDFs, GitHub repositories, and YouTube videos to compare papers, extract methods, or audit manuscripts. If a study design or output format is ambiguous, Kady can pause and show a short in-chat question form - including multiple choice and free text - rather than guessing.
The software includes 149 scientific skills covering genomics, proteomics, bioinformatics, drug discovery, chemistry, materials science, and clinical research. It also offers 326 guided workflow templates across 22 disciplines that turn common analyses into fill-in-the-blank starting points. A Living Lab Notebook records the reasoning trail, logging hypotheses, methods, observations, decisions, and linked artifacts. Users can connect external tools through MCP, a plug-in standard for AI assistants, to add reference managers, GitHub, databases, and other services.
Model and compute options
The app itself is free, but users are responsible for their own provider charges and limits. Kady supports several connection methods. Users can connect supported subscriptions - including ChatGPT Plus/Pro, Claude Pro/Max, GitHub Copilot, and xAI - directly through Pi OAuth. They can use major hosted models from OpenAI, Anthropic, Google, xAI, Qwen, and others through a single OpenRouter account. Direct API keys are supported for Anthropic, OpenAI, Google Gemini, DeepSeek, Mistral, Groq, and more than a dozen other providers. NVIDIA NIM models can be used with an API key from build.nvidia.com, billed against NVIDIA API credits.
For work that must not leave the machine, users can run free local models through Ollama or any OpenAI-compatible server such as LM Studio or vLLM. Kady tracks OpenRouter usage and Anthropic OAuth metered usage against project spending caps. ChatGPT, Copilot, and xAI subscription usage is tracked separately because those providers manage their own quotas and overages.
Demanding computation can be moved to Modal, which provides on-demand cloud CPU or GPU environments. Kady persists and monitors the job, stages validated inputs, brings outputs back into the local project, and reserves estimated compute cost against the project budget. Long jobs survive chat turns and backend restarts.
Benchmark performance
On the CompBioBench leaderboard, K-Dense BYOK ran the same harness four times through OpenRouter, each run using one model and one isolated project per question. GPT-6 Astra with xhigh reasoning reached 98% accuracy for the #2 spot. Gemini 3.8 Flash with xhigh reasoning scored 94%, GPT-5.6 SOL with xhigh reasoning scored 92%, and DeepSeek V4 Pro with high reasoning scored 81%. Ranks will move as new submissions arrive.
K-Dense also published an internal benchmark comparing K-Dense BYOK, Claude Science, and Biomni Lab across scientific quality and research execution in a 20-prompt test. K-Dense BYOK was configured to use Claude Opus 4.8 with the xHigh reasoning level. The company noted that the benchmark "was designed, run, and evaluated internally by K-Dense rather than an independent third party, so the results should be interpreted as an internal evaluation under the tested setup, not as a universal measure of platform performance."
Getting started
K-Dense BYOK runs on macOS, Linux, or Windows 10/11. On Windows, users need Node.js 22+ and Git for Windows first - Kady's agent runs shell commands through the Git Bash that Git for Windows provides. Installation requires cloning the repository from GitHub and running a start script. The first start installs everything automatically, then opens a browser to localhost:3000. The app is currently in beta.
Why this matters for researchers and scientists
K-Dense BYOK puts a full research assistant on a scientist's own computer without requiring coding experience or handing data to a third-party platform. The combination of local file storage, flexible model selection - from free local models to paid cloud providers - and built-in scientific skills means researchers can choose the cost and privacy level that fits each project. The CompBioBench results suggest the tool can handle real computational-biology tasks at high accuracy when paired with a capable model. For researchers who need end-to-end workflows with managed infrastructure and team collaboration, K-Dense also offers K-Dense Web, a hosted platform built for professional and academic research teams.
Your membership also unlocks:
