Dante

Dante is a chat interface for local and remote LLMs, focused on an elegant, distraction-free UX.

Features

• OpenAI-compatible and Anthropic models — point Dante at any OpenAI-compatible endpoint (e.g. OpenCode Go) or an Anthropic-style API.
• Streaming + reasoning — answers stream token by token, with the model's thinking shown in a collapsible block.
• Runs live on the server — a reply keeps generating when you reload or navigate away, and you can stop it at any time. Partial answers are saved after every step.
• Synced across devices — every open tab or device gets live updates over Server-Sent Events, including replies that were started elsewhere.
• Conversation history — chats are stored in SQLite and listed in the sidebar, grouped by date (Heute, Gestern, …), and can be renamed or deleted.
• Rich messages — Markdown with syntax-highlighted code blocks, plus copy, edit and retry actions.
• Elegant UX — shadcn/ui (Rhea style) with light/dark/system theming, a resizable sidebar and a composer with Enter to send and Shift+Enter for a newline.

Status

Dante is an early prototype. The model is configured through environment variables, and the settings page is still a mock.

Tech stack

• Frontend: React 19, Vite, shadcn/ui on Base UI, Tailwind CSS 4, Vercel AI SDK (@ai-sdk/react), TanStack Query, React Router
• Backend: Node.js 24, Hono, Vercel AI SDK (ai, @ai-sdk/openai-compatible, @ai-sdk/anthropic)
• Database: SQLite via the built-in node:sqlite (server/dante.db, created on first run)

Getting started

Requires Node.js 24+, pnpm and Task.

# Install dependencies and create server/.env from the example
task setup

# Fill in the model settings in server/.env, then start the web app (Vite on
# :5173) and the server (Hono on :3000) with live reload
task dev

Open http://localhost:5173 and send a message to see it stream.

Without Task, run pnpm install, copy server/.env.example to server/.env, and start pnpm dev in the root and in server/.

Configuration

The server reads server/.env:

Variable	Description
PORT	Server port (default: 3000)
DANTE_DB_PATH	SQLite database file (default: dante.db)
DANTE_MODEL_TYPE	openai-compatible (default) or anthropic
DANTE_MODEL_URL	Base URL of the model API, e.g. https://opencode.ai/zen/go/v1
DANTE_MODEL_ID	Model to use, e.g. deepseek-v4-flash
DANTE_MODEL_API_KEY	API key for the model provider
DANTE_API_TOKEN	Bearer token for /api/* routes; leave empty to disable auth

The web app doesn't send a token yet, so keep DANTE_API_TOKEN empty for now.

Development

task typecheck   # web app and server
task lint        # oxlint
task format      # Prettier
task check       # all of the above, e.g. before committing

To add shadcn/ui components, run pnpm dlx shadcn@latest add <component>; they land in src/components/ui.

Roadmap

• MCP servers — connect tool servers (stdio or HTTP) so the agent can use tools like Filesystem or GitHub.
• Agentic loops — multi-step tool use with the tool calls shown in the chat.
• Settings — configure providers, models and assistants in the UI instead of env vars.