K-Dense BYOK
A local research workspace where Kady executes scientific analysis, evidence review, and report generation from plain-language requests.
What does this agent do, and when should you use it?
K-Dense BYOK is an open-source desktop research application centered on Kady, an AI research assistant for scientists. In a project workspace, users describe work in plain language and Kady can inspect files, write and run analysis code, search and read sources, and produce figures, reports, and an execution record. Projects, conversations, notebooks, and results remain in ordinary local folders; hosted-model request material goes directly to the provider selected by the user, while Ollama and other OpenAI-compatible servers support local models. The application includes projects, parallel chat tabs, a Living Lab Notebook, scientific file previews, LaTeX editing, and optional Modal compute, making it a supervised research-execution environment rather than a chat-only assistant.
After a user creates a project and adds or attaches files, Kady can clean data, check quality, select and run statistical methods, compare groups, fit models, and create publication-ready figures. It can search the web and read web pages, PDFs, GitHub repositories, and YouTube videos, while also accepting text, data, and image inputs. Its work leaves scripts, intermediate files, tables, figures, reports, and Living Lab Notebook entries in the project; the notebook can export Markdown, JSON, an artifact bundle, or PDF and can generate a manuscript-style Methods draft. Models can be connected in Settings → Model providers through openai-codex, anthropic, GitHub Copilot, or xAI subscriptions, through OpenRouter, or through Ollama / an OpenAI-compatible server; MCP adds external research tools, and Modal provides on-demand cloud CPU or GPU compute.
- A bioinformatics researcher has counts.csv and needs a treated-versus-control differential-expression analysis with a volcano plot.
- A scientist needs to search and read papers, web pages, PDFs, GitHub repositories, or YouTube material before producing a methods or evidence summary.
- A lab member wants analysis code, generated figures, decisions, and linked artifacts captured in an exportable Living Lab Notebook.
- A researcher handling sensitive data wants local project storage and a local model through Ollama or another OpenAI-compatible server.
- A researcher with a long-running workload wants to select a Modal cloud CPU or single-/multi-GPU environment for a chat and bring outputs back into the local project.
What are this agent's strengths and limitations?
- Local project folders retain scripts, artifacts, conversations, and results for inspection, backup, movement, or use with other software.
- It joins data analysis, literature review, scientific file previewing, figure creation, and reporting in one research-execution flow rather than offering text responses alone.
- The same interface supports OpenRouter, several subscription sign-ins, Ollama, and other OpenAI-compatible servers, with model choice per chat.
- The Living Lab Notebook can connect hypotheses, methods, observations, decisions, code, and artifacts, then export the record in multiple forms.
- The product is currently in beta, and the documentation states that further features and improvements are expected.
- Free application access does not make model or compute usage free: OpenRouter, Anthropic OAuth metered extra usage, and optional Modal compute may incur charges and are subject to project budget controls.
- Material needed for a hosted-model request is sent to the selected provider; strict data-locality requirements call for local models instead.
- Windows requires Node.js 22+ and Git for Windows, and the first start takes several minutes to install dependencies.
How do you install or deploy this agent?
You need macOS, Linux, or Windows 10/11 and at least one model source: an OpenRouter API key, a supported ChatGPT Plus/Pro, Claude Pro/Max, GitHub Copilot, or xAI subscription, or local Ollama models. On macOS/Linux:git clone https://github.com/K-Dense-AI/k-dense-byok.gitcd k-dense-byokcp .env.example .env./start.sh
On Windows, first install Node.js 22+ and Git for Windows, then run:git clone https://github.com/K-Dense-AI/k-dense-byok.gitcd k-dense-byokcopy .env.example .env.\start.cmd
The first start installs dependencies automatically and opens http://localhost:3000; .env is optional and can hold an OpenRouter key.
How do you use this agent?
Open the local application, create a project, add data or attach project files in a chat, then issue a specific request such as: “Run a differential expression analysis on counts.csv comparing treated vs control, and plot a volcano plot.” For subscriptions, sign in under Settings → Model providers and choose a model and reasoning level; for hosted pay-as-you-go access, configure an OpenRouter API key. Inspect streamed tool calls and generated files while work runs, add up to five follow-up messages to a running conversation, or stop it; review, annotate, or export the resulting Living Lab Notebook.
How does this agent compare with similar options?
The README presents a K-Dense-designed, run, and evaluated internal 20-prompt benchmark comparing K-Dense BYOK with Claude Science and Biomni Lab. K-Dense BYOK used Claude Opus 4.8 with xHigh reasoning for those runs; the results are not an independent evaluation or a universal platform ranking.