A customizable, open-source frontend for the Venice AI API.

Venice gives you access to powerful AI models for text, images, audio, music, and video. OpenVenice gives you a clean interface to use them — one you own, can modify, and can host yourself.

Your interface, your rules. Venice's official UI is great, but sometimes you want more control:

• Customize everything — add tools, change layouts, tweak parameters, build features that matter to you. The codebase is intentionally simple and hackable.
• Share your API key with family — host OpenVenice on your own server, enter your key once, and give your family a clean AI interface without them needing their own accounts.
• No server, no backend — it's a static site. Your API key stays in your browser and goes directly to Venice's API. Nothing passes through a middleman.
• Transparent — every API call is visible in the source. No telemetry, no analytics, no tracking.
• Barebones on purpose — ships with useful features like visual workflows, but keeps things minimal so you can build on top without fighting existing complexity.

Streaming responses, conversation history, model selection, web search, citations, temperature control. Markdown rendering with safe URL handling and syntax highlighting. Image attachments (paste, drag-drop, or upload) for vision models.

Prompts, negative prompts, style presets, steps, resolution, aspect ratio, variants, watermark control. Lightbox gallery with download. Edit / Upscale / Background-remove tools in a separate tab.

Text-to-speech with 50+ voices across 9 languages. Adjustable speed, multiple formats (MP3, Opus, AAC, FLAC, WAV). Audio transcription via Whisper.

Text-to-music with optional lyrics, duration control, instrumental mode. Supports Stable Audio, ACE-Step, ElevenLabs, MiniMax, MMAudio.

Text-to-video and image-to-video. Configurable aspect ratio, resolution, duration. Auto-detects model capabilities. Cancellable polling with elapsed-time display.

Vector embeddings for text with selectable models and dimension display.

Visual node editor for chaining models. Connect Input → LLM → Image Gen → Output, with full parameter controls per node. Independent branches run in parallel. Starter templates for album covers, podcast episodes, music videos, song writing, character portraits, and story scenes.

Conversational agent that builds and edits workflows on a live canvas as you describe them in plain language.

npm install
npm run dev

Open http://localhost:5173, click API Key in the header, paste your Venice AI API key, and start using it.

By default, your key is held in sessionStorage — meaning it's gone when you close the tab. If you check Remember across sessions, the key is encrypted with a passphrase you choose (AES-GCM via PBKDF2, 250k iterations, all in-browser) and stored in localStorage. Your passphrase is never persisted; you re-enter it on each new session.

You can disconnect at any time from the API key dialog.

You have three options. Pick whichever fits.

Build the app and serve /dist from any static host (Netlify, Vercel, Cloudflare Pages, S3, Pi, …). The browser will hit https://api.venice.ai directly. CORS must allow your origin — Venice's API does.

npm run build       # outputs to /dist

Override the API base URL at build time if you proxy elsewhere:

VITE_VENICE_BASE_URL=https://my-proxy.example.com/api/v1 npm run build

A multi-stage Dockerfile is included; runtime is Nginx serving the built bundle with a SPA fallback.

docker build -t openvenice .
docker run -p 8080:80 openvenice

A railway.json is included for one-click Railway deploy.

If you'd rather keep API calls server-side (so the key never touches a browser), front the static bundle with any reverse proxy that forwards /api/v1/* to https://api.venice.ai/api/v1/* and strips/injects the key. The frontend already speaks to a relative /api/v1 base when VITE_VENICE_BASE_URL is unset in dev (it uses the Vite proxy automatically).

Deploy it, share the URL, have them enter your API key (or their own). Each browser stores its own key and history.

React 19, TypeScript, Vite, Zustand, TanStack Query, Tailwind CSS v4, React Flow.

src/
├── app.tsx                     # Tab routing + error boundary
├── components/
│   ├── chat/                   # Chat interface + message bubbles
│   ├── image/                  # Image generation + edit/upscale/bg-remove
│   ├── audio/                  # TTS + transcription
│   ├── music/                  # Music generation
│   ├── video/                  # Video generation (cancellable polling)
│   ├── embeddings/             # Embeddings
│   ├── workflows/              # Visual workflow editor
│   ├── playground/             # Agent-authored workflows
│   ├── layout/                 # Sidebar, header, API key dialog
│   └── ui/                     # Shared components, ErrorBoundary, Toaster
├── stores/                     # Zustand stores (versioned + quota-safe)
├── hooks/                      # Data hooks, useBlobUrl, useChat
├── lib/
│   ├── venice-client.ts        # fetch w/ retry+backoff
│   ├── stream.ts               # robust SSE parser
│   ├── workflow-engine.ts      # parallel topological executor
│   ├── workflow-validator.ts   # graph + param validation
│   ├── workflow-mutations.ts   # patch reducer + auto-layout
│   ├── playground-agent.ts     # validated patch parser
│   └── safe-storage.ts         # localStorage with quota recovery
└── types/                      # TypeScript types

• Retry + backoff on transient API failures (429/5xx), respecting Retry-After.
• Cancellable streams and polling — Stop button kills any in-flight request.
• Capped polling with elapsed-time display so video/music jobs can't run forever.
• Object-URL lifecycle managed via useBlobUrl — no leaked blob references.
• CSP in index.html restricts script/style/connect sources.
• Error boundary at the root + per-tab so one bad render doesn't take the whole app down.
• Toast system for user-facing errors (network failures, transient issues).
• Persisted state versioned + quota-safe — schema migrations, automatic pruning on QuotaExceededError.
• Markdown sanitization — javascript: and other unsafe URI schemes are stripped from rendered AI output.
• Accessible — every interactive control is a real <button>, focus-visible rings everywhere, aria labels on icon-only controls, prefers-reduced-motion honored.

Fork it, break it, make it yours. PRs welcome.

MIT