K-Dense BYOK: a desktop AI research assistant that runs its own code

The gap K-Dense BYOK is trying to fill

Most chat assistants answer a scientific question and stop. The README draws the line explicitly: Kady is designed to carry out research work, not only answer questions. That means the output of a session is not a paragraph but a set of files, scripts, tables, figures, and a notebook recording what happened. The stated audience is scientists in any field, with no coding experience required, since the user describes the task and Kady writes and runs the code. The BYOK part of the name is the other half of the pitch. You bring an OpenRouter key, a supported AI subscription, or a local Ollama model, and the app itself stays free. Projects, conversations, notebooks, and results live in ordinary folders on your machine rather than on K-Dense infrastructure. That combination, an agent that executes code plus local storage plus user-supplied model access, is a specific position. It is not a hosted notebook and not a plain chat client.

How Kady actually runs an analysis

The mechanism visible in the material is a tool-calling loop over a project folder. You describe a task, Kady activates relevant skills from the 149 it ships with, and it inspects files, writes and runs code, searches and reads sources, and produces figures and reports. The README states that the scripts, intermediate files, tables, figures, and reports remain available for inspection and reuse, which is the part that matters: nothing is trapped in a proprietary result object. Input is not limited to typed text. You can dictate a request, upload files through the project browser, attach project files to a conversation, or paste and drop images into a message for a vision-capable model. When a requirement is ambiguous, the documentation says Kady can pause and present a short in-chat question form with multiple choice, free text, and image input instead of guessing. Steering is asynchronous: up to five follow-up messages can be added to a running conversation, and a run can be stopped at any time. Skills are listed in Settings and can be disabled, which is the only documented control over what the agent reaches for.

Bring your own keys, and bring your own spending discipline

Model access is the configuration surface, and the README is unusually direct about the consequences. You can use prepaid OpenRouter, connect a supported AI subscription, or run free local models via the Ollama path documented in docs/local-models-ollama.md. The app tracks paid OpenRouter usage and Anthropic OAuth's documented metered extra usage against project spending caps. ChatGPT, Copilot, and xAI subscription usage is tracked separately because those providers manage quotas and overages, and the README states plainly that a subscription login does not imply unlimited or free usage. That is a real constraint, not a footnote. An agent that writes and runs code can consume tokens in bursts that a chat session never would, and only some of the supported providers expose a cap the app can enforce. The privacy boundary follows the same logic. When a hosted model is used, the material needed for that request goes directly to the provider under that provider's terms. The README's own recommendation is to use a local Ollama model when data must not leave the machine, which is the honest framing: local-first storage is not the same as local-only inference.

Getting it running: what the documentation actually specifies

The README does not print an install command line. It points to a YouTube walkthrough, Getting Started with K-Dense BYOK, covering installation and a first research task, and the repository is a TypeScript project with the server package under server/, since the version badge links to server/package.json. Setup therefore starts from the release page or the walkthrough rather than from a copy-pasteable curl line, and that is a fair criticism of the documentation as it stands. What is specified concretely is the post-install configuration: choose a provider (OpenRouter, a supported subscription, or Ollama), set project spending caps where the provider supports metering, and optionally add search-provider keys, since web search works without an additional account but optional keys improve capacity. Skills are browsed and disabled in Settings. The local model path has its own document at docs/local-models-ollama.md. If you are evaluating the project, the walkthrough plus that file are the two things to read before installing anything.

Where the design stops short

The project describes itself as beta, and several documented behaviours read as beta-shaped. Skill activation is described as automatic, with Settings offering only browse and disable, so there is no documented way to pin a skill or force a procedure for a given run. The skill, workflow, and database counts are presented as coverage, but coverage counts say nothing about whether a particular method is implemented correctly, and the README does not claim otherwise. The internal benchmark figure compares K-Dense BYOK against Claude Science and Biomni Lab on a 20-prompt benchmark, with K-Dense BYOK configured to use Claude Opus 4.8 at the xHigh reasoning level. The README itself notes the benchmark was designed, run, and evaluated internally rather than by an independent third party, so it should be read as an internal evaluation under one tested setup. Treat it as a description of the test conditions, not as evidence about your workload. The wrong-tool case is any setting where the analysis must be reproducible by someone else on a different machine without the app, or where an audited, versioned pipeline is required. An agent that writes code on demand is a convenience, not a validated instrument.

How it differs from Biomni Lab

Biomni Lab appears in the README's own benchmark comparison and is the closest point of reference in the material. The difference in approach is where the work happens and who supplies the model. K-Dense BYOK is a desktop application with a local project folder, and the model comes from keys or subscriptions you already hold, with Ollama as the option that keeps inference on the machine. That arrangement puts the storage boundary and the billing boundary in your hands, and it puts the maintenance burden there too: provider credentials, spending caps, and model choice are yours to manage. The README positions the app against hosted alternatives generally, noting that K-Dense does not host or store your workspace. If your constraint is that data cannot leave your machine, the Ollama path is the deciding feature. If your constraint is that you want someone else to run the infrastructure and the model, this project is pointed the other way.

Licence, maintenance, and upgrade cost

The repository is MIT licensed, which permits commercial and private use and modification, subject to the usual requirement to preserve the copyright and permission notice. That is a permissive licence and it is stated in the repository metadata and the README badge. Nothing in the material suggests a separate commercial tier for the app itself, and the README says the app is free with provider charges and limits remaining yours. On maintenance, the release cadence is visible in the metadata: v0.10.0, v0.9.19, and v0.9.18 all landed within roughly a day of each other, and the README notes many features and improvements are on the way. That pace cuts both ways. You get fixes quickly, and you should expect the surface to move under you. The version badge in the README still reads 0.7.3 while the latest release is v0.10.0, which is a small but concrete sign that the documentation trails the code. If you pin a version, budget time for re-reading Settings and the local-models document after upgrades.

Editorial conclusion

Adopt K-Dense BYOK if your analysis already lives in files on your own machine and you want an agent that writes scripts you can read, rerun, and archive, with a local Ollama model available for work that cannot leave the machine. Skip it if you need a validated instrument, an audited pipeline, or a hosted multi-user environment, because the project labels itself beta and its only published comparison was designed, run, and evaluated internally by K-Dense. Before committing, verify three things: that your provider is on the supported list, that your project folder is on a disk you are willing to let an agent write to, and that a single end-to-end run on your own dataset produces scripts you can actually rerun outside the app.

Official sources

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