K-Dense BYOK

K-Dense BYOK (Bring Your Own Keys) is an open-source AI research assistant named Kady designed for scientists. It allows users to describe research tasks in plain language (e.g., analyze this dataset, review my manuscript) and executes them within a local research workspace.

Key characteristics:

• No coding required: Kady writes and runs the necessary analysis code automatically.
• Local-first privacy: Projects, conversations, and notebooks are stored in ordinary folders on your local machine. K-Dense does not host or store your data.
• Flexible Compute: You can use your own AI subscriptions (ChatGPT, Claude, etc.), hosted API providers (OpenRouter), or free local models (Ollama) to power the assistant.

Kady's agent has access to a powerful local shell to install packages and run analyses. This shell runs as your current operating-system user.

Critical Security Notes:

• No OS-level isolation: The shell is not a security boundary. A process running as you can read your own .env files, ~/.kady directory, and other local secrets.
• Adversarial Files: Do not ask Kady to process untrusted files that may contain prompt injections designed to access your local secrets.
• Recommended Isolation: Use an OS sandbox, container, VM, or a separate user account when working with untrusted content.
• Skill Security: Installing a skill from a third-party source is equivalent to running code from that source. Review third-party skills before installation and prefer pinning to a specific branch or tag.

Kady uses a tabbed interface for chats with the following constraints:

• Tab Limit: There is a hard cap of 10 tabs per project. To open a new tab once this limit is reached, you must close an existing one.
• Persistence: Browser refreshes and tab closures preserve workspaces, chat tabs, drafts, queues, and live turns. However, stopping or restarting the Kady backend will end in-flight turns (though completed history remains on disk).
• Workflow Launching: Workflows always launch into the currently active tab. If you have a long-running task in Tab A and want a new workflow to run in Tab B, you must switch to Tab B before clicking 'Launch'.

K-Dense BYOK uses independent research workspaces called Projects.

• Isolation: Each project has its own files, chats, notebooks, model choices, and spending policies.
• Parallelism: You can run multiple projects simultaneously. Within a single project, you can use up to 10 parallel chat tabs, each with its own conversation, model, and cost, while sharing the same project files.
• Persistence: The workspace state (open files, panel sizes, active turns) is restored after browser refreshes or closures. A live turn will reconnect to the running backend.
• Local Storage: All work is stored in ordinary local files, making projects easy to back up, move, or archive.

Kady supports up to 10 parallel chat tabs within a single project:

• Creation: Click + in the chat tab strip to open a new chat.
• Isolation vs. Sharing: Each tab maintains its own message history, model choice, attached files, and cost meter. However, all tabs share the project's files, meaning files produced in one chat are immediately available in others.
• Background Processing: Tabs continue working in the background even when you switch to a different tab.
• Management: Double-click a tab title to rename it. Closing a tab cancels any active work in that tab.

Kady supports three main ways to access AI models:

1. OpenRouter: Provides pay-as-you-go access to models from OpenAI, Anthropic, Google, etc. Requires an API key starting with sk-or-....
2. Pi OAuth subscriptions: Connect existing subscriptions for ChatGPT Plus/Pro (openai-codex), Claude Pro/Max (anthropic), GitHub Copilot (github-copilot), or xAI (xai). To connect, go to Settings $\rightarrow$ Model providers in the app and click Connect.
3. Local Ollama: Run free local models without hosted-provider credentials.

Important: OpenRouter-only features (like OpenRouter Fusion and server-side speech transcription) require an OpenRouter API key even if you have a Pi subscription.

To prevent overspending, K-Dense reserves a worst-case estimate before creating Modal resources using the formula:

estimated hourly rate × requested timeout

Admission Control: New jobs are blocked if settled project spend + open reservations + new reservation exceeds the hard project cap.

Cost UI Terms:

• spent: Settled model and compute estimates.
• reserved: Worst-case holds for active Modal jobs.
• committed: The sum of spent and reserved.

Provenance tracks whether the current bytes on disk match the bytes that existed when a tool produced them. This prevents citations in lab notebooks from silently pointing to incorrect data after a file has been regenerated.

Staleness Statuses:

• Current: The bytes on disk match the producing step's recorded hash.
• Stale: The file has changed since the step that produced it.
• Unverified: No hash is available to compare against (e.g., due to size limits or harvest timing); sameness is not confirmed.