An AI-powered multi-agent classroom that turns any topic into an interactive lesson.

English · 中文

---

What is this?

Nova is a multi-agent teaching platform. Type a topic on the home page, and AI breaks it into a structured course outline, builds slides, writes narration, and delivers the lesson in a virtual classroom where multiple AI agents work together: a lead teacher, a teaching assistant, and a class clown who keeps the mood light.

It is not just a slide generator. It is a full teaching system with role separation, safety guardrails, and knowledge tracing.

Home page: enter any topic, or open the cached demo to try a full classroom instantly

Demo Video

~10 min walkthrough: real LLM generation flow (topic → outline → agents → slides → classroom) + full classroom interaction (playback / 24-scene browse / 14 settings tabs).

Screenshots

Home & topic input	LLM generation	Classroom playback

Scene sidebar	Interactive lab	Quiz scene

Settings panel	Token usage	Model config

Quick Start

Prerequisites

• Node.js 22+
• pnpm 10+

Install

git clone https://gitcode.com/badhope/nova.git
cd nova
pnpm install

Slow install? See docs/DEPENDENCY-TIERING.md for tiered installs:

• Just want to run it fast: pnpm install --prod (skips dev toolchain, ~1–2 min)
• Maintainers, install extras on demand: node scripts/install-tiers.mjs core|extras|all

Missing optional features (charts, math, code highlight, MCP, document parsing, …) never block startup — they degrade gracefully (plain text / placeholder / guided error). Open Settings → Capability Status to see what's installed vs. missing and copy the install command in one click.

Configure

Create a .env.local with at least one LLM provider:

# Option A: direct API key
OPENAI_API_KEY=your-key
OPENAI_BASE_URL=https://api.openai.com/v1

# Option B: server-side managed config (recommended — keys stay server-side)
cp server-providers.yml.example server-providers.yml
# Edit server-providers.yml with your credentials

Optional environment variables:

DEFAULT_MODEL=openai:your-model        # Default model (provider:model format)
LLM_TIMEOUT_MS=300000                  # LLM request timeout in milliseconds
FALLBACK_MODELS=openai:m2,openai:m3    # Comma-separated fallback models
LLM_THINKING_DISABLED=true             # Disable thinking/reasoning tokens
SKIP_TS_CHECK=true                     # Skip TypeScript checking during build

Run

pnpm dev

Open http://localhost:3000 and enter a topic to start.

No API key? Try the demo

Click "Open Cached Demo Course" on the home page to load a pre-built Introduction to AI course — no API key required.

Features

Course Generation

• AI Outline Generation — Breaks a topic into progressive scenes ordered by knowledge dependencies
• Slide Production — Each scene ships with titles, bullet points, and flow diagrams
• Voice Narration — The AI teacher narrates each scene with natural TTS across multiple engines
• Interactive Quizzes — Auto-generated multiple-choice and fill-in-the-blank questions with real-time scoring
• Knowledge Graph — Visual concept maps that connect key ideas across the course
• PBL Mode — Project-based learning with interactive practice tasks

Multi-Agent Classroom

Classroom: a virtual room where AI teacher, assistant, and class clown co-teach

Agent	Role	Permissions
AI Teacher	Leads the lesson and explains core concepts	Speak, slide control, spotlight, whiteboard
AI Assistant	Supports the teacher and answers questions	Speak, whiteboard, slide control
Class Clown	Lightens the mood	Speak

• Role Persistence — Customize names, descriptions, and permissions for 10 built-in roles; changes persist across sessions
• Runtime Constraints — Per-role max_actions and max_turns enforced at runtime
• Discussion Orchestration — A Director Graph manages turn-taking and discussion flow

Prompt Engineering & Governance

• 34 Templates — Covering outline generation, content creation, action sequencing, and quiz generation
• Snippet System — Role guidelines and action types stored as Markdown snippets, editable without recompiling
• Guardrails — PII detection, toxicity filtering, and hallucination scanning on every generated scene
• Skill Registry — 5 registered skills gated by a whitelist
• REST API — GET /api/prompts to list templates, GET /api/skills to list skills

Infrastructure

17 LLM Providers

Provider	Example Models
OpenAI	GPT-5.6, GPT-5.4
Azure OpenAI	User-defined deployments
Anthropic	Claude Opus 4.8, Claude Sonnet 4.6
Google	Gemini 3.5 Flash, Gemini 2.5 Pro
DeepSeek	DeepSeek-V4-Pro, DeepSeek-V4-Flash
Qwen	Qwen3.7 Plus, Qwen3.6 Flash
GLM	GLM-5.2, GLM-4.6
Kimi	Kimi-K2.7, Kimi-K2.6
MiniMax	MiniMax-M3
SiliconFlow	Full model aggregation
Doubao	Doubao Seed series
OpenRouter	DeepSeek, and more
Grok	Grok 4.20, Grok 4.1
Tencent Hunyuan	Hy3 Preview
Xiaomi MiMo	MiMo V2.5 Pro, MiMo V2
Ollama	Local models
Lemonade	Local AMD models

• TTS — OpenAI, SiliconFlow, Doubao, Minimax, Volcano
• Image Generation — SiliconFlow, Minimax, ComfyUI
• Web Search — Tavily, SearXNG
• Document Parsing — AliDocMind, MinerU
• MCP Tools — Connect external tools via Model Context Protocol
• i18n — English, Simplified Chinese, Traditional Chinese, Japanese, Korean, Arabic, Portuguese, Russian
• Dark Mode — Site-wide support

Enterprise Features

• Quota Management — Per-user generation quotas with QUOTA_EXCEEDED (402) responses
• Input Validation — All generation API routes validate input length, topic, and requirements
• Audit Logging — All API actions logged with retention policies (90-day default)
• Rate Limiting — Configurable per-endpoint rate limits
• SSRF Protection — URL allowlist/denylist for outbound requests
• Content Moderation — PII detection, toxicity filtering, and hallucination scanning
• Role-Based Access — 10 configurable agent roles with fine-grained permissions
• Knowledge Tracing — Bayesian knowledge tracing for student progress tracking

Architecture

Data flow: user enters a topic → the prompt engine assembles the prompt → LLM generates content → guardrails scan for safety → multi-agent orchestration → interactive classroom rendering. State is persisted to browser-local storage via Zustand.

Testing

pnpm test          # Unit & component tests (340 files / 3155 cases)
pnpm test:e2e      # End-to-end tests (Playwright)
pnpm test:e2e:ui   # E2E with interactive UI
pnpm lint          # ESLint
pnpm typecheck     # TypeScript type checking

E2E tests cover the full flow: home → generation → classroom navigation → quiz interaction. All tests use mock APIs — no LLM key needed.

Project Structure

nova/
├── app/                  # Next.js App Router
│   ├── api/              # API routes (prompts, skills, generate/*)
│   └── [locale]/         # i18n routing
├── lib/                  # Core logic
│   ├── ai/               # Multi-LLM provider integration
│   ├── agent/            # Multi-agent runtime
│   ├── choreography/     # Animations & effects
│   ├── guardrails/       # Safety pipeline
│   ├── orchestration/    # Role management & constraints
│   └── prompts/          # Prompt templates & snippets
├── components/           # React components
├── packages/             # Workspace sub-packages
│   └── @nova/
│       ├── dsl/          # Domain type definitions
│       ├── renderer/     # Slide rendering engine
│       ├── importer/     # Document import
│       └── storage/      # Persistence layer
├── e2e/                  # Playwright tests
├── configs/              # Shared constants
└── assets/               # Static assets & screenshots

Tech Stack

Layer	Technology
Framework	Next.js 16 (App Router, Turbopack)
Language	TypeScript 5.9
UI	React 19, Tailwind CSS 4, Radix UI
State	Zustand (persisted)
AI	Vercel AI SDK, multi-provider
Testing	Vitest, Playwright
Package Manager	pnpm Workspaces

Mirrors

Platform	URL	Role
GitCode (primary)	https://gitcode.com/badhope/nova	Canonical source, issues & PRs
GitHub (mirror)	https://github.com/weed33834/nova	Read-only mirror
Gitee (mirror)	https://gitee.com/badhope/nova	Read-only mirror

GitCode is the primary development platform. Please open issues and pull requests on GitCode. The three mirrors are kept in sync on every push.

FAQ

Q: The "Start" button on the home page is disabled (grey)? A: No usable LLM provider is configured. Check OPENAI_API_KEY in .env.local, or configure server-side credentials in server-providers.yml (recommended — keys stay server-side).

Q: Generation fails with "model returned no valid scene content"? A: Usually the model's quota is exhausted or temporarily unavailable. Run node scripts/verify-models.mjs to probe availability, then set a working model in DEFAULT_MODEL / LLM_FALLBACK_MODELS. Verified working: qwen3.8-max (fast) and glm-5.2 (stable).

Q: pnpm dev is slow on classroom routes / ChunkLoadError? A: Dev mode compiles routes on demand (tens of seconds on first visit). For demos/recordings, use pnpm build && pnpm start (production mode — all routes precompiled).

Q: How do I enable the "Edit Course" (Pro mode) button? A: Set NEXT_PUBLIC_NOVA_EDITOR_ENABLED=true in .env.local and rebuild. This variable is inlined into client bundles at build time — runtime-only changes won't take effect.

Q: Other startup issues? A: See docs/OPERATIONS-HANDBOOK.md — a troubleshooting handbook with 17 historical issues (symptom → root cause → fix → prevention) and a pre-flight checklist.

Contributing

Issues and pull requests are welcome. Please read CONTRIBUTING.md before submitting.

License

MIT