KIN

Local-first workspace for AI chats, projects, and agent workflows. Includes local application state and optional model providers;

Comment

KIN

A local-first workspace for AI chats and agent workflows.

KIN brings conversations, projects, agent definitions, tasks, model settings, and automation controls into one desktop-oriented workbench. The repository contains a TypeScript core, a React interface, a Tauri/Rust desktop shell, and SQLite-backed local state.

KIN is under active development. Some paths are connected in source but have not been validated end to end. This README describes the current project without treating planned behavior as a certainty.

Audit and current limitations · Architecture notes · API · Skills · Slash commands · FAQ · Contributing · License

What KIN includes

  • Chat and workspace concepts: projects, channels, direct messages, and agent identities.
  • Agent workflows: a default orchestrator, specialist definitions, goals, tasks, and dependency relationships.
  • Model connections: credential settings and a model gateway with local Ollama and several hosted-provider adapters.
  • Automation and extensions: schedules, recurring routines, skills, and MCP connections.
  • Local application state: SQLite and project files are stored on the machine running KIN.
  • Desktop and browser integration: source paths exist for browser and computer-control actions; their security and packaged runtime behavior still need further validation.

These are product areas present in the codebase, not a claim that every flow is complete or production-ready. See the dated audit for confirmed gaps and what has or has not been demonstrated.

Run from source

Requirements

  • Node.js 20 or later
  • npm
  • Rust and the platform build tools only if you are working on the Tauri desktop shell
  • Ollama only if you want to try a local model

Install and build

git clone https://github.com/abhayzangir1/KIN.git
cd KIN
npm install
npm run build

Start the core daemon in one terminal:

npm run daemon --workspace=core

Start the web workbench in another terminal:

npm run dev --workspace=ui

Open http://localhost:5173. The core listens on 127.0.0.1:54321; it is bound to loopback and is not exposed as a LAN server by default.

On Windows, start.bat and on macOS/Linux, start.sh are convenience development launchers. They require a working Node installation and repository dependencies; they are not self-contained installers.

To run the core test suite yourself:

npm test --workspace=core

No test count or passing status is stated here because results depend on the checkout and environment.

Models and credentials

KIN has settings and APIs for provider credentials, model discovery, and agent model assignment. The current catalog still contains fallback entries, and readiness indicators do not reliably prove that a key is valid or that a model can be called. OpenRouter free/paid filtering and price information should be treated as incomplete. Verify a provider with an actual request before relying on it.

If you configure a hosted provider, prompts and related context sent for inference leave your machine and are handled by that provider under its own terms. Browser actions send requests to the websites you visit. MCP servers receive the data passed to their tools. Ollama can run local inference when it is installed, running, and has the selected model; this does not by itself prove that every part of the application is offline.

Data and security

KIN stores application state in a local SQLite database and uses the local filesystem for project data. Local storage does not mean every operation stays on-device: hosted model calls, browser navigation, and connected MCP tools can communicate with external services.

The source includes capability checks, approval flows, secret handling, task leases, and event recording. These controls have gaps documented in the current audit. In particular, task evidence can be accepted without verification, the manual task-status endpoint can synthesize a verified sign-off, and some browser/computer execution is not isolated by a host process sandbox. Do not interpret the presence of these controls as a security certification.

Current limitations

  • A current runtime pass with a signed-in browser and real provider keys has not been completed.
  • Model readiness can be overstated when Ollama is unavailable or a hosted API key is invalid.
  • Some task-completion paths can report completion without valid acceptance evidence.
  • Startup actively admits and dispatches queued runs across platform restarts.
  • Git worktrees provide separate working directories; they are not system-level process or network sandboxes.
  • Clean-machine desktop packaging has not been demonstrated in this audit.

The audit report distinguishes source findings, historical runtime observations, and unverified behavior. Product and technical specifications in docs/PRD.md and docs/TRD.md describe target requirements, not an immediate production feature promise.

Screenshots

The images below show interface views included in the repository. They are visual references and do not demonstrate that the associated workflow is currently verified.

KIN workbench

KIN agent topology view

KIN settings and credentials view

KIN agent inspector

Releases

See GitHub Releases for any published artifacts. Available files and supported platforms are release-specific. A clean-machine installation and bundled Node/core runtime were not verified for this audit.

Project documents

The tutorials and example skills are workflow sketches. Their presence does not establish that each example is a live integration or verified end-to-end recipe.

License

KIN is distributed under the MIT License.

Built With

Share this project:

Updates

Submission history