Local-storage AI chat with a built-in agent workflow engine

What it does

LLMChat.co is a monorepo housing a web and desktop chat application built on Next.js. It wraps multiple LLM providers into a single interface and offers specialized research modes like Pro Search and Deep Research. All chat history lives in the browser’s IndexedDB via Dexie.js, so conversations stay on your device.

The interesting bit

Instead of leaning entirely on external frameworks, the project ships a custom workflow orchestration engine with typed event emitters, shared context, and task routing. You can chain steps—planner, gatherer, analyzer, reporter—into a directed graph that streams progress back to the UI in real time.

Key highlights

• Supports six LLM backends: OpenAI, Anthropic, Google, Fireworks, Together AI, and xAI.
• Monorepo layout splits concerns across apps (web, desktop) and packages (ai, orchestrator, ui, etc.).
• Research agents break queries into sub-tasks, synthesize findings, and emit structured reports.
• State management uses Zustand; the UI layer leans on Shadcn UI, Tailwind, and Tiptap.
• Data storage is fully client-side, though LLM calls still hit remote APIs.

Caveats

• The architecture diagram includes a prisma package for database schema, which the README does not reconcile with the “no server-side storage” privacy claim.
• The workflow examples import the raw openai SDK rather than the AI SDK listed in the tech stack, so the exact abstraction boundary is unclear.

Verdict

Worth a look if you want a reference implementation for client-side AI storage or a typed workflow engine in TypeScript. Skip it if you need a mature, batteries-included platform with clear server-side sync.