Auto PPT

Open-source PowerPoint generation engine for AI agents. Give it a prompt, optional source files, and an LLM key — get back a validated .pptx.

Quick Start

# 1. Install
npm install && python -m pip install -r requirements.txt

# 2. Configure (writes .env with your LLM key)
./auto-ppt init

# 3. Generate — mock mode, no LLM key needed
./auto-ppt generate --mock \
  --prompt "Create an 8-slide AI strategy deck for executives" \
  --source examples/inputs/sample-source-brief.md

# 4. Generate — real model, after init
./auto-ppt generate \
  --prompt "Create an 8-slide AI strategy deck for executives" \
  --source examples/inputs/sample-source-brief.md

# 5. Revise
./auto-ppt revise \
  --deck output/py-generated-deck.json \
  --prompt "Compress to 6 slides, make it more conclusion-driven"

Output: output/py-generated-deck.json + .pptx, output/py-revised-deck.json + .pptx.

What It Does

Capability	Detail
Prompt-to-deck planning	Natural-language prompt → structured slide deck
Natural-language revision	Iterate on an existing deck with free-text instructions
Source ingestion	.txt .md .csv .json .yaml .xml .html .pdf .docx, images, URLs
Schema validation	JSON-schema check before every render
LLM providers	OpenAI, OpenRouter, Claude, Gemini, Qwen, DeepSeek, GLM, MiniMax, any OpenAI-compatible endpoint
PPTX rendering	JS renderer (pptxgenjs) or Python renderer (python-pptx with brand templates)
Security	Path traversal protection, SSRF blocking, file size limits, subprocess timeout

Entry Points

Interface	Command	Use Case
CLI	./auto-ppt generate / revise	Interactive or scripted usage
MCP	python mcp_server.py	Claude Desktop, Cursor, Windsurf
HTTP	python py-skill-server.py	REST integration (POST /skill)
JSON skill	python py-agent-skill.py --request req.json	File-based agent orchestration
Docker	docker compose up --build	One-command deploy

MCP Configuration

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "auto-ppt": {
      "command": "python",
      "args": ["/absolute/path/to/auto-ppt-prototype/mcp_server.py"]
    }
  }
}

Cursor / Windsurf — add to .cursor/mcp.json or .windsurf/mcp.json:

{
  "mcpServers": {
    "auto-ppt": {
      "command": "python",
      "args": ["/absolute/path/to/auto-ppt-prototype/mcp_server.py"]
    }
  }
}

Tools exposed: create_deck, revise_deck. Both accept sources, mock mode, and optional output_dir.

Docker

export OPENAI_API_KEY="sk-..."
docker compose up --build                          # HTTP skill server
docker run --rm -it -e OPENAI_API_KEY auto-ppt-prototype python mcp_server.py  # MCP stdio

Testing

python -m pytest tests/ -v   # 300+ tests
npm run ci:smoke             # JS renderer + end-to-end smoke

CI: pytest on Python 3.10 / 3.11 / 3.12, smoke on Node.js 18 / 20 / 22.

Architecture

prompt + sources ➜ Python smart layer ➜ deck JSON ➜ PPTX renderer ➜ .pptx
                        ↑ revise loop ↲

• Python (python_backend/): planning, revision, source loading, LLM calls, schema validation
• Node (generate-ppt.js): pptxgenjs rendering from validated deck JSON
• deck JSON: the stable contract between both layers

Documentation

Doc	Content
Examples	Copy-paste usage flows
User Guide	Day-to-day usage
Integration Guide	HTTP, MCP, JSON skill patterns
Product Overview	Positioning and boundaries
Repository Map	File-by-file structure
Changelog	Version history
Roadmap	Planned evolution

Multilingual: all guides available in English, 中文, 日本語.

License

MIT