lokulAI desktop app

The product problem

lokulAI started with a practical constraint: an AI workspace should be useful with local data and local models, without hiding the state of work behind a generic chat box.

The product combines project-scoped conversations, documents, tasks, agents, schedules, memory, voice, and learning tools. Keeping those surfaces coherent meant treating context, persistence, and operator control as core product architecture.

The desktop architecture

I built the workspace with Next.js, React, TypeScript, and SQLite, then packaged it in Electron for macOS arm64 and Windows x64. Ollama is the local model path, with optional provider abstractions where a workflow needs them. Desktop data lives in the operating system user-data directory rather than the source checkout.

The desktop shell adds native behavior such as tray operation, notifications, voice capture, file drops, watched folders, and update checks. Project-scoped document ingestion supports local extraction and citation-ready retrieval across text, PDF, DOCX, CSV, and JSON sources.

Building proof into the release

Desktop-only API routes use a per-launch loopback token, and exported diagnostics redact secrets, prompts, content, and local paths. A fresh-profile smoke exercises first-run auth and onboarding, settings, watched-folder updates, voice draft capture, learned-skill execution, and route protection.

The release gate also runs clean installation, native-binding checks, the core test suite, type checks, lint, browser smoke, desktop tests, and installer creation. Separate hosted-surface evidence verifies the public auth and update-feed contract while keeping the private workspace routes off the public web surface.

The takeaway

lokulAI is an example of local-first AI as a full product discipline: desktop packaging, durable context, supervised workflows, privacy boundaries, and release evidence all have to work together.