|
|
|
K-Dense Updates
// July 24, 2026
|
|
Plenty of AI tools can produce a confident scientific answer. Far fewer can show you how they got there, or help you go find out whether it’s actually true. That gap is the thread running through everything below.
Our new partnership with Litmus Science takes a hypothesis out of the browser and into a physical lab. Our head-to-head against Claude Science found our clearest advantage is producing work another scientist can audit and rerun. And this week’s two BYOK releases are largely about making research inspectable.
|
>From Hypothesis to Wet Lab
K-Dense Web can search 250+ scientific databases, analyze your data, and design the experiment that would settle a question. What it cannot do is pick up a pipette. Our new partnership with Litmus Science is about closing that half of the loop.
Litmus Science turns a well-specified experimental question into quotes from vetted labs – both robotics-run autonomous facilities and traditional CROs – usually within about a week, so you can compare cost, timeline, and fit before committing to anything. We’re now working to bring that directly into K-Dense Web, along with an MCP server that will let the agent scope and submit experiments programmatically.
Read the partnership announcement →
|
>K-Dense Web vs. Claude Science: 20 Tasks
We ran both platforms through the same 20 scientific research tasks, scoring not just whether the answer was right but whether the work could be traced, reproduced, and rerun by someone else.
K-Dense Web led 14 of the 20 prompts and won research execution on all 20, averaging 14.7/25 against 6.4/25. The widest gaps were in whether an analysis could actually be regenerated and traced back to its sources – the difference between a result you can read and one you can rerun.
We designed, ran, and judged this benchmark ourselves, so the writeup includes the full methodology, the complete prompt-level results, and the limitations.
See where each platform excelled →
|
>BYOK v0.9.0: A Nature Medicine Reproduction
K-Dense BYOK is our free, open-source co-scientist. It runs on your own desktop, powered by the accounts and API keys you choose, and your projects, notebooks, and results stay in ordinary folders on your machine.
To show what it can actually do, we used it to reproduce a recent Nature Medicine publication end to end – the whole run, start to finish.
Watch the walkthrough →
Two releases shipped this week, v0.8.0 and v0.9.0. What they change for you at the bench:
|
•Use the AI subscription you already pay for. Connect Claude Pro/Max, ChatGPT Plus/Pro, GitHub Copilot, or xAI directly in Settings instead of setting up pay-as-you-go API billing – or run everything on free local models through Ollama when data cannot leave the building. Claude Opus 5 is the new default.
|
|
•Long computations now survive. Send heavy work to an on-demand cloud CPU or GPU and it keeps running in the background through chat turns and restarts, with outputs written back into your project and estimated cost reserved against the project budget before anything starts.
|
|
•Results arrive as structured cards. Statistical tests, QC checks, dataset schemas, tables, plots, citations, and molecules render as typed cards linked to the underlying files, so you can open the actual artifact behind any number.
|
|
•Specialists have to show their evidence. The 21 scientific specialists now work under contracts that require evidence-based reporting, and the activity view names which specialist produced which piece of the work.
|
|
•Annotate papers as you read them. Live expert annotations on PDFs inside the app, alongside the analysis they inform.
|
Get K-Dense BYOK on GitHub →
|
>177 Research Sessions You Can Inspect
Wondering what a full research session actually looks like? Our use case library spans science, finance and investment, engineering and technology, health and climate, and geopolitics and strategy.
Recent additions range from mapping tau pathology pathways and auditing anti-tau therapeutics, to a techno-economic analysis of cultivated meat scale-up, to simulating the long-term orbital stability of the Beta Pictoris debris disk. Most include the full session and a downloadable PDF report, so you can inspect exactly how the work was done.
Browse the use case library →
|
>Build Your Own Agent Skills: Workshop Recording
Missed our agent skills workshop earlier this month? The recording is a beginner-friendly walkthrough: installing a ready-made skill, then building one from scratch, live. Worth an hour if you’ve been meaning to automate a piece of your own workflow.
We especially appreciated everyone who stayed after the deck to show us their skills in practice – great work.
Watch the full recording →
|
|
|
|
One question before you go: if you could hand a single experiment to a lab this month, what would it be? Hit reply and tell us – we’re using those answers to decide what the Litmus Science integration should support first.
Start a research session →
|
|
Until next time,
The K-Dense Team
|
|
|
K-Dense Inc.
380 Portage Avenue
Palo Alto, CA 94306
|
Unsubscribe
|
|
|