K-Dense BYOK: a local-first AI co-scientist you run with your own model keys

The problem K-Dense BYOK addresses: research work, not chat answers

Most AI assistants answer questions. K-Dense BYOK, whose assistant is named Kady, is built around a different expectation: that the model should carry out the work and leave artifacts behind. The README describes tasks such as analyzing a dataset, reviewing a manuscript, searching the literature, and building a figure, with Kady inspecting files, writing and running analysis code, and keeping a record of what it did.

The audience is explicit in the README: scientists in any field, with no coding experience required. The repository topics point the same way, listing bioinformatics, drug discovery, scientific computing, and research assistant alongside agentic-ai and local-first. The bundled material is sized for that audience: the README badges claim 149 scientific skills, 326 guided workflow templates across 22 disciplines, and 229 databases.

That combination is the interesting part. A general coding agent can already run Python in a folder. What K-Dense BYOK adds is a curated layer of scientific procedures and a workspace that treats scripts, intermediate files, tables, figures, and reports as first-class outputs rather than console noise.

How Kady works: workspace on disk, model calls to the provider you pick

The architecture visible in the repository is a Node application with a server and a web front end. The top level holds server/, web/, scripts/, start.mjs, start.sh, start.cmd, and an .env.example. There is no compiled binary and no container file at the root. The package.json is private and exposes two scripts, start and check, both routed through start.mjs.

The data flow follows from the README's local-first claim. Projects, conversations, notebooks, and results live in ordinary folders on your machine, and K-Dense states it does not host or store them. When a request goes to a hosted model, the material needed for that request is sent directly to the provider you selected, under that provider's privacy terms. Running a local Ollama model is the documented way to keep data on the machine.

Model access is the BYOK part. The .env.example treats OpenRouter as the default route to hosted models, with OPENROUTER_API_KEY able to stay blank if every chat uses a connected Pi subscription or local Ollama. It also notes that OpenRouter Fusion and the server-side speech transcription fallback remain OpenRouter-only, so a fully local or subscription-only setup has documented gaps. Because the provider is an OpenAI-compatible endpoint, OPENROUTER_BASE_URL can point at another gateway while reusing OPENROUTER_API_KEY for that gateway's key. NVIDIA NIM is a separate path: the example file says usage draws on NVIDIA-managed credits rather than per-token dollar pricing, so it is recorded but never counted against a project spend cap.

Spend tracking is deliberately uneven. The README states that Kady tracks paid OpenRouter usage and Anthropic OAuth's documented metered extra usage against project spending caps, while ChatGPT, Copilot, and xAI subscription usage is tracked separately because those providers manage quotas and overages. The README adds that a subscription login does not imply unlimited or free usage. Anyone budgeting a lab should read that sentence twice.

Installing K-Dense BYOK and running a first task

The repository does not ship an installer package. Installation is a clone plus the start script, and the engines field in package.json requires Node 22 or newer. The root package.json defines the two entry points:

{
  "name": "k-dense-byok",
  "private": true,
  "scripts": {
    "start": "node start.mjs",
    "check": "node start.mjs --check"
  },
  "engines": {
    "node": ">=22"
  }
}

From a checkout, run the check target first. It validates the environment before the app starts, which is cheaper than debugging a failed launch:

npm run check
npm start

Before starting, copy .env.example to .env and fill in the provider you intend to use. OpenRouter is the documented default for hosted models; the key can stay blank only if every chat uses a connected Pi subscription or local Ollama. If you point the app at another OpenAI-compatible gateway, set the base URL as well:

OPENROUTER_API_KEY=your-key-here
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1

The README also points to a separate document, docs/local-models-ollama.md, for running free local models, and to Settings, API keys for pasting provider credentials live. Once the app is up, the first real use is the one the README demonstrates: describe a task in plain language, such as analyzing a dataset or reviewing a manuscript, and watch the tool calls stream while Kady writes and runs code in the project. Expect files to appear in the workspace as it goes. Kady can also pause and show an in-chat question form, with multiple choice, free text, and image input, when a study design or output format is ambiguous.

Where K-Dense BYOK stops being the right tool

The local-first claim has a boundary, and the README states it plainly: when you use a hosted AI model, the material needed for that request goes to the provider you selected. Local-first here means the workspace stays on your machine, not that the analysis never leaves it. For data that cannot be sent to a third party, the documented path is a local Ollama model, and the README does not claim feature parity between the two. The .env.example names two features that stay OpenRouter-only, so a local-only configuration is a documented subset.

Cost control is partial by design. Project spending caps cover paid OpenRouter usage and Anthropic OAuth's metered extra usage. Subscription usage on ChatGPT, Copilot, and xAI is tracked separately because those providers manage quotas and overages. If your budget depends on a hard ceiling across every provider, the caps do not give you one.

The project also labels itself beta in the README, and the release list is dense rather than calm: three releases on 2026-09-08 alone, including v0.10.0. Frequent releases are not a defect, but they do mean interfaces and settings can move between versions, and any workflow you build should be pinned to a version you have checked.

One more caveat sits in the README itself. The benchmark figure comparing K-Dense BYOK with Claude Science and Biomni Lab across 20 prompts was designed, run, and evaluated internally, with K-Dense BYOK configured to use Claude Opus 4.8 at the xHigh reasoning level. The README says the results should be read as an internal evaluation under that setup, not a universal measure. That is the right framing, and it is the only framing the repository supports.

K-Dense BYOK compared with a general coding agent

The closest alternative is a general-purpose coding agent such as Claude Code pointed at a project folder. The difference is in what ships in the box. A coding agent gives you a shell, a file editor, and a model; you supply the domain layer yourself. K-Dense BYOK bundles the domain layer: 149 scientific skills, 326 workflow templates across 22 disciplines, and a library of databases, with skills that Kady activates automatically and that you can browse or disable in Settings.

The trade-off runs the other way too. A coding agent is not tied to a scientific skill catalogue, so it does not need to be updated when that catalogue changes, and its behaviour is not shaped by template conventions you did not choose. K-Dense BYOK also adds a specific interface for scientific work, including the in-chat question form for ambiguous study designs and image input for vision-capable models. If your work is mostly software engineering with occasional data analysis, the bundled scientific layer is weight you will not use.

The README's own comparison point is Claude Science, which it names in the benchmark figure and in the title of the project description. K-Dense positions itself as the local-first, bring-your-own-keys option in that space. The repository does not include a feature-by-feature comparison against Claude Science, so the honest summary is that K-Dense BYOK differs in deployment model and model choice, and its internal benchmark is a K-Dense-run evaluation.

Licence, maintenance and the cost of upgrading

The licence is MIT, per the repository badge and the LICENSE file at the root. MIT permits commercial and private use with attribution and without warranty. That is a permissive, low-friction licence, and it also means the project offers no support commitment. Nothing here is legal advice; if you are deploying inside a regulated environment, the licence text and your own obligations are what matter.

Maintenance looks current rather than dormant. The repository is not archived, the last push was on 2026-09-08, and the release list shows v0.10.0 on that same date, following v0.9.19 and v0.9.18 earlier the same day. Three releases in a day is the pattern of a project moving quickly, not one coasting.

The upgrade cost is the flip side of that pace. The README describes a settings surface with skills, specialists, API keys, and spending caps, and the .env.example documents provider variables that can also be set live in the UI. A fast release cadence means those surfaces can shift. The practical approach is to pin to a release tag, read the release notes before moving, and keep your .env under version control separately from the checkout so a pull does not overwrite your provider configuration. The project's own version badge in the README reads 0.7.3 while the newest release is v0.10.0, which is a small but real reminder that documentation and code drift.

Editorial conclusion

K-Dense BYOK fits scientists who want an agent that writes and runs analysis code inside a workspace they control, and who are willing to supply their own OpenRouter key, subscription, or local Ollama model. It does not fit anyone who needs an air-gapped pipeline by default, since hosted models receive the material needed for each request, nor teams that want a hosted, managed service. Before adopting it, check the Node version against the engines field, confirm which provider keys you can supply, and read the .env.example comments about which features remain OpenRouter-only.

Frequently asked questions

Official sources

1. K-Dense-AI/k-dense-byok on GitHub
2. License: MIT
3. Project website
4. README
5. Releases