Open-source AI co-scientist runs on your own desktop
The MIT-licensed app ships 149 science skills, 326 workflow templates and 21 sub-agents, and expects you to plug in your own model keys.
Symbolic image: a research workstation with glowing status LEDs runs while someone, seen from behind, plugs in a cable.
K-Dense BYOK is an MIT-licensed desktop app that runs research tasks locally and drives them with model keys you supply yourself rather than a bundled model.
At a glance
- Version 0.7.3, marked beta, MIT license; needs Node.js 22 or newer on macOS, Linux or Windows 10/11.
- 149 science skills, 326 workflow templates across 22 disciplines and 21 specialized sub-agents.
- 229 data resources across 18 categories; more than 60 scientific file formats open in preview.
- Models arrive via OpenRouter, Nvidia NIM, existing subscriptions or locally through Ollama — none ship with the app.
- The comparison against Claude Science rests on 20 in-house prompts, not on independent testing.
K-Dense BYOK is an open-source desktop application that does research work on your own machine and deliberately ships without a language model: you bring the credentials.
What the project page claims
The assistant is called Kady and takes instructions in ordinary language. According to the repository it analyzes datasets, produces publication-ready figures, works through PDFs, web pages, code repositories and videos, and writes up the result. A running lab notebook records the intermediate reasoning.
The counts the project puts forward: 149 science skills covering genomics, chemistry, drug discovery and clinical research, 326 workflow templates across 22 disciplines, 229 data resources in 18 categories, and 21 sub-agents for narrower assignments. More than 60 file formats open in preview, and LaTeX documents can be edited in place.
You supply the key
BYOK is short for bring your own key. Nothing is billed by the project itself; it connects to OpenRouter, to Nvidia NIM models, to subscriptions you already hold at ChatGPT, Claude, GitHub Copilot or xAI, and to local models served through Ollama. Run it locally and no third party gets paid at all.
Setup means cloning the repository and running a start script, after which the interface opens in a browser at localhost:3000. Node.js 22 or newer is required.
The benchmark is the vendor's own
The repository carries an internal comparison against Claude Science and Biomni Lab over 20 prompts, run on Claude Opus 4.8 at the xHigh reasoning setting. The authors state plainly that this is their own assessment under particular conditions rather than a general platform ranking. The Claude Science but better line in the subtitle is therefore a self-description, not a third-party measurement.
What stays unverified
Exactly one source was available for this piece: the project page itself. That leaves the feature list, the benchmark figures and the practical value of the skills in real lab work unconfirmed. The popularity numbers — roughly 1,200 stars and 146 forks — are a snapshot of the repository page, not evidence of how widely the tool is actually used.
FAQ
What is K-Dense BYOK?
An open-source desktop application under the MIT license that runs a research assistant called Kady on your own machine, using model access you provide yourself.
Which models does K-Dense BYOK work with?
Per the project page: OpenRouter, Nvidia NIM models, existing ChatGPT, Claude, GitHub Copilot and xAI subscriptions, plus local models via Ollama or an OpenAI-compatible server.
Is K-Dense BYOK free to use?
The software is, under the MIT license. Costs come from whichever model access you connect; running locally through Ollama means paying no third party.
