paper-demo-agent 0.3.0

AI agent that reads any scientific paper and builds a live demo

Paper Demo Agent

Turn any scientific paper into a live interactive demo — with one command.

Quick Start · Features · Output Formats · Providers · CLI Reference · Python API

---

Quick Start

No API key needed — authenticate with Claude Code or Gemini CLI:

pip install paper-demo-agent

# Option A: Use Claude Pro/Max subscription (free — no API billing)
claude login                    # one-time: sign into claude.ai
paper-demo-agent demo 1706.03762

# Option B: Use Gemini CLI (free tier: 60 req/min, 1000 req/day)
gemini                          # one-time: sign into Google
paper-demo-agent demo 1706.03762 --provider gemini

# Option C: Traditional API key
paper-demo-agent key set ANTHROPIC_API_KEY sk-ant-...
paper-demo-agent demo 1706.03762

Or launch the web UI:

paper-demo-agent ui
# → Opens http://localhost:7860

No pip needed (with pipx):

pipx run paper-demo-agent ui

System Dependencies (optional)

Most output formats are pure Python/HTML and need no extra setup. For Graphviz diagrams (--form diagram --subtype graphviz), the system-level Graphviz binary is required:

# macOS
brew install graphviz

# Ubuntu / Debian
sudo apt-get install graphviz

For LaTeX/Beamer output (--form presentation --subtype beamer), a TeX distribution is needed:

# macOS
brew install --cask mactex-no-gui

# Ubuntu / Debian
sudo apt-get install texlive-latex-recommended texlive-fonts-extra

---

Features

	Feature	Description
🧠	Smart Routing	AI reads the paper and picks the best demo type — model inference, data explorer, slides, diagram, and more
📦	10 Output Formats	Gradio apps, Streamlit dashboards, reveal.js slides, PowerPoint, LaTeX/Beamer, project pages, blog articles, READMEs, Mermaid flowcharts, Graphviz diagrams
🤖	Multi-Phase Agent	4-phase pipeline (Research → Build → Polish → Validate) with tool use, prior-work analysis, and self-correction
🔌	6 LLM Providers	Anthropic, OpenAI, Google Gemini, DeepSeek, Qwen, MiniMax — switch with one flag
🔑	Zero-Config Auth	Auto-detects credentials from Claude Code (OAuth), Gemini CLI (OAuth), OpenAI Codex CLI, Aider, and gcloud ADC — no API keys needed
📄	Any Input	arXiv ID, arXiv URL, any URL, local PDF, or raw text
🌐	Web UI	Dark-themed Gradio interface with real-time progress streaming, phase stepper, file preview, and one-click demo launch
⚡	One Command	paper-demo-agent demo 1706.03762 — that's it

---

How It Works

flowchart LR
    A[Paper Input] --> B[Parse]
    B --> C[Analyze]
    C --> D{Route to Skill}
    D --> E[Research Phase]
    E --> F[Build Phase]
    F --> G[Polish Phase]
    G --> H[Validate Phase]
    H --> I[Live Demo]

    style A fill:#6366f1,stroke:#4f46e5,color:#fff
    style I fill:#22c55e,stroke:#16a34a,color:#fff

1. Parse — Fetches the paper from arXiv, URL, or local PDF. Extracts title, abstract, full text, and figures.
2. Analyze — AI classifies the paper type (model, dataset, algorithm, framework, survey, theory, empirical) and recommends the best output format.
3. Route — Maps paper type + user preferences to one of 15 specialized skills.
4. Research — Finds the paper's official resources (GitHub, HuggingFace, project page) and identifies foundational prior work to give the build agent a bigger picture. Library docs are pre-baked in the skill prompts.
5. Build — Multi-iteration code generation with 7 tools (write_file, read_file, list_files, web_search, run_python, extract_pdf_page, execute_command).
6. Polish — Quality review pass with skill-specific checklists.
7. Validate — Form-compliance check ensures the output matches the requested format, with auto-correction if needed.

---

Output Formats

Apps

Format	Technology	Best For
Gradio App	Gradio 5 (Python)	Model inference, interactive demos, HuggingFace Spaces
Streamlit App	Streamlit (Python)	Data dashboards, exploration tools, widgets

Presentations

Format	Technology	Best For
HTML Slides	reveal.js 5.2.1	Animated talks with KaTeX math and speaker notes
PowerPoint	python-pptx 1.0.0	Conference presentations, offline sharing
LaTeX / Beamer	Beamer + Metropolis	Academic talks with TikZ diagrams and booktabs

Pages

Format	Technology	Best For
Project Page	HTML/CSS/JS	Nerfies/Distill.pub-style research landing pages
Blog Article	HTML + D3.js + KaTeX	Interactive explainers with visualizations
GitHub README	Markdown + Mermaid	Publication-quality README with badges and diagrams

Diagrams

Format	Technology	Best For
Mermaid Flowchart	Mermaid.js v11 (ESM)	Interactive architecture diagrams, step-by-step walkthroughs
Graphviz Diagram	Python graphviz	Publication-quality SVG/PNG architecture diagrams

Gallery

Interactive Diagram	reveal.js Presentation	Project Website	Gradio App
ResNet — Mermaid flowchart	Attention — HTML slides	Attention — project page	BERT — Gradio demo

Made with Paper Demo Agent — community showcase

Paper	arXiv	Format	Command
Attention Is All You Need	1706.03762	Presentation	paper-demo-agent demo 1706.03762 --form presentation
Attention Is All You Need	1706.03762	Project Page	paper-demo-agent demo 1706.03762 --form page
Deep Residual Learning	1512.03385	Flowchart	paper-demo-agent demo 1512.03385 --form diagram
BERT	1810.04805	Gradio App	paper-demo-agent demo 1810.04805 --form app

Want to add yours? Open an issue with the title [Gallery] Paper Name — format and a screenshot!

---

Supported Providers

Provider	Default Model	Env Variable	Notes
Anthropic	claude-sonnet-4-6	ANTHROPIC_API_KEY	Best quality. Auto-detected from Claude Code
OpenAI	gpt-5.2	OPENAI_API_KEY	Also supports o3, o4-mini. Auto-detected from Codex CLI
Gemini	gemini-2.5-flash	GOOGLE_API_KEY	Auto-detected from Gemini CLI or gcloud ADC
DeepSeek	deepseek-chat	DEEPSEEK_API_KEY	Also supports deepseek-reasoner
Qwen	qwen-max	QWEN_API_KEY	Also supports qwen-plus, qwen-turbo
MiniMax	abab6.5-chat	MINIMAX_API_KEY	Also requires MINIMAX_GROUP_ID

Zero-Config Authentication

Paper Demo Agent automatically detects credentials from CLI tools you already use — no API keys needed:

Tool	Auth Method	Provides	Setup
Claude Code	OAuth (macOS Keychain / ~/.claude/)	ANTHROPIC_API_KEY	claude login — uses your Pro/Max subscription
Gemini CLI	OAuth (Google Cloud Code Assist)	GOOGLE_API_KEY	gemini — free tier: 60 req/min, 1000 req/day
OpenAI Codex CLI	~/.codex/auth.json	OPENAI_API_KEY	Auto-detected
Aider	~/.aider.conf.yml	ANTHROPIC_API_KEY, OPENAI_API_KEY	Auto-detected
gcloud ADC	Application Default Credentials	Gemini auth	gcloud auth application-default login

Priority order: Claude Code / Gemini CLI → saved config → environment variables → Codex / Aider.

If you use any of these tools, Paper Demo Agent works immediately — no key setup required.

💡 Recommended: Install Claude Code (npm i -g @anthropic-ai/claude-code && claude login) or Gemini CLI (npm i -g @google/gemini-cli && gemini) for the fastest zero-config experience.

---

CLI Reference

# Generate a demo (AI picks the best format)
paper-demo-agent demo 1706.03762

# Specify provider and model
paper-demo-agent demo arxiv:1706.03762 \
    --provider anthropic \
    --model claude-opus-4-6

# Specify output format
paper-demo-agent demo 1706.03762 --form presentation
paper-demo-agent demo 1706.03762 --form app --subtype streamlit
paper-demo-agent demo 1706.03762 --form presentation --subtype beamer
paper-demo-agent demo 1706.03762 --form page --subtype readme
paper-demo-agent demo 1706.03762 --form diagram --subtype graphviz

# From a local PDF
paper-demo-agent demo paper.pdf --provider openai

# Custom output directory
paper-demo-agent demo 1706.03762 --output ./my-demo

# Launch the web UI
paper-demo-agent ui
paper-demo-agent ui --port 8080 --share

# Manage API keys
paper-demo-agent key set ANTHROPIC_API_KEY sk-ant-...
paper-demo-agent key set OPENAI_API_KEY sk-...
paper-demo-agent key list

# List available providers
paper-demo-agent providers

# HuggingFace (for gated models)
paper-demo-agent login   # Browser-based HF login
paper-demo-agent logout

---

Python API

from paper_demo_agent import PaperDemoAgent

agent = PaperDemoAgent(provider="anthropic")

# Full pipeline: parse → analyze → generate
result = agent.run(
    source="1706.03762",
    demo_form="app",           # or: presentation, slides, latex, website, flowchart, ...
    demo_type="user_demo",     # or: findings, theoretical
    on_progress=print,
)

print(f"Output:  {result.output_dir}")
print(f"Main:    {result.main_file}")
print(f"Run:     {result.run_command}")

Step-by-Step

# Parse only
paper = agent.parse("1706.03762")
print(f"Title: {paper.title}")
print(f"Abstract: {paper.abstract[:200]}...")

# Analyze
analysis = agent.analyze(paper)
print(f"Paper type: {analysis.paper_type}")
print(f"Recommended form: {analysis.demo_form}")
print(f"Recommended type: {analysis.demo_type}")

---

Architecture

graph TB
    subgraph Input
        A1[arXiv ID/URL]
        A2[PDF Upload]
        A3[Text/URL]
    end

    subgraph Core
        B[PaperDemoAgent]
        C[Paper Parser]
        D[Paper Analyzer]
        E[Skill Router]
    end

    subgraph Skills
        S1[ModelInferenceSkill]
        S2[DataExplorerSkill]
        S3[AlgorithmVisualizerSkill]
        S4[FrameworkTutorialSkill]
        S5[SurveyDashboardSkill]
        S6[TheoreticalExplainerSkill]
        S7[FindingsDashboardSkill]
        S8[FlowchartGeneratorSkill]
        S9[GraphvizDiagramSkill]
        S10[ReadmeGeneratorSkill]
        S11[BlogExplainerSkill]
        S12[StreamlitDemoSkill]
        S13[GeneralQASkill]
    end

    subgraph Generator["Multi-Phase Generator"]
        G1[Research Phase]
        G2[Build Phase]
        G3[Polish Phase]
        G4[Validate Phase]
    end

    subgraph Providers
        P1[Anthropic]
        P2[OpenAI]
        P3[Gemini]
        P4[DeepSeek]
        P5[Qwen]
        P6[MiniMax]
    end

    A1 & A2 & A3 --> B
    B --> C --> D --> E
    E --> S1 & S2 & S3 & S4 & S5 & S6 & S7 & S8 & S9 & S10 & S11 & S12 & S13
    S1 & S2 & S3 & S4 & S5 & S6 & S7 & S8 & S9 & S10 & S11 & S12 & S13 --> G1 --> G2 --> G3 --> G4
    G1 & G2 & G3 & G4 -.-> P1 & P2 & P3 & P4 & P5 & P6

    style B fill:#6366f1,stroke:#4f46e5,color:#fff
    style G2 fill:#6366f1,stroke:#4f46e5,color:#fff

Project Structure

paper_demo_agent/
├── agent.py              # Main orchestrator (parse → analyze → route → generate)
├── config.py             # ~/.paper-demo-agent/config.json key storage
├── cli.py                # CLI entry point (click)
├── ui/app.py             # Gradio web interface
├── providers/            # LLM providers (Anthropic, OpenAI, Gemini, DeepSeek, Qwen, MiniMax)
├── paper/                # Paper parsing (arXiv API, PDF extraction, URL fetching)
├── analysis/             # Paper classification + routing
├── skills/               # 15 specialized demo generation skills
│   ├── model_inference.py
│   ├── data_explorer.py
│   ├── algorithm_visualizer.py
│   ├── framework_tutorial.py
│   ├── survey_dashboard.py
│   ├── theoretical_explainer.py
│   ├── findings_dashboard.py
│   ├── flowchart_generator.py
│   ├── graphviz_diagram.py
│   ├── readme_generator.py
│   ├── blog_explainer.py
│   ├── streamlit_demo.py
│   ├── general_qa.py
│   └── ...
└── generation/           # Multi-phase generator + tool definitions
    ├── generator.py      # 4-phase pipeline (Research → Build → Polish → Validate) + prior-work analysis
    ├── tools.py          # 7 agent tools (write_file, read_file, web_search, ...)
    └── runner.py         # Demo launcher (open in browser / run app server)

---

Development

git clone https://github.com/wshuai190/PaperDemoAgent
cd paper-demo-agent
pip install -e ".[dev]"
pytest tests/

Contributing

Contributions are welcome! See CONTRIBUTING.md for the full guide.

• Bug reports — include the arXiv ID and provider used
• New skills — subclass BaseSkill and register in the router
• New providers — subclass BaseLLMProvider and add to factory.py
• Gallery submissions — generated a great demo? Open an issue with a screenshot!

---

License

MIT — see LICENSE

---

If Paper Demo Agent saved you time, give it a star!