Auto PPT Engine

Generate and revise editable PowerPoint decks from prompts, source files, or agent tool calls.

Auto PPT Engine is a self-hostable backend for teams that want AI agents, internal tools, or automation workflows to produce .pptx output through MCP, CLI, HTTP, or Docker.

Status: beta-quality open-source backend. Not a SaaS. Not a polished end-user GUI. Built for agent workflows, internal automation, and custom integrations. Current scope is feature-complete enough to evaluate today: prompt in, validated deck JSON and .pptx out.

Why use it?

• Turn a prompt or a source brief into an editable .pptx deck
• Let an AI agent create or revise slides through MCP, HTTP, or a JSON skill entrypoint
• Keep generation in your own environment with self-hosted deployment options
• Add engineering guardrails: schema validation, chart fallback, visual QA, and theme-aware rendering

Who is this for?

This is the backend behind “generate me a PPT”. It is designed to be embedded into your workflow, not used as a standalone SaaS product.

Quick Start

Prerequisites: Python 3.10+ and Node.js 18+.

Fastest path: mock mode, no API key required

# 1. Install dependencies
npm install && pip install .

# 2. Generate a deck locally without any LLM key
./auto-ppt generate --mock \
  --prompt "Create an 8-slide AI strategy deck for executives" \
  --source examples/inputs/sample-source-brief.md

Output:

• output/py-generated-deck.json
• output/py-generated-deck.pptx

Real model generation

# 1. Configure your LLM key into .env
./auto-ppt init

# 2. Generate with a real model
./auto-ppt generate \
  --prompt "Create an 8-slide AI strategy deck for executives" \
  --source examples/inputs/sample-source-brief.md

Revise and validate

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

# Run visual QA on the rendered PPTX
./auto-ppt qa-visual output/py-generated-deck.pptx --strict

qa-visual writes a JSON report (default: alongside PPTX in <deck-name>-qa/visual-qa-report.json) and attempts to export slide images when soffice and pdftoppm are available.

Example Run

$ ./auto-ppt generate --mock --prompt "Q1 AI Strategy Review for Leadership"

Action: create
Deck JSON: output/py-generated-deck.json
PPTX:      output/py-generated-deck.pptx
Renderer:  pptxgenjs
Slides:    8
Sources:   1

  1. [title       ] Q1 AI Strategy Review for Leadership
  2. [agenda      ] Agenda
  3. [bullet      ] Background and goals
  4. [two-column  ] Current state and challenges
  5. [process     ] Python smart layer workflow
  6. [timeline    ] Execution timeline
  7. [chart       ] Adoption metrics
  8. [closing     ] Key recommendations

The output is not just a .pptx file. The system also emits a validated deck JSON contract that can be revised, re-rendered, audited, or handed to another workflow step.

Example Output

These thumbnails come from real .pptx files already checked into the repository under output/.

What they show:

• A prompt-driven first draft deck
• A layout coverage deck used to exercise more slide types
• A revised deck produced from an existing JSON contract

Prompt-driven deck	Layout coverage deck	Revised deck
output/py-agent-generated-deck.pptx	output/new-layouts-business-clean.pptx	output/py-agent-revised-deck.pptx
First-pass deck generated by the Python smart layer	Coverage deck exercising newer layout types	Revision workflow output from an existing deck JSON

Theme Gallery

The 6 built-in themes below are shown using thumbnails regenerated from the current renderer. Each thumbnail is the same deck rendered with a different built-in theme, so the visual differences are easy to compare.

business-clean	corporate-blue	dark-executive
Clean default for general business decks	More formal blue-toned presentation style	Dark boardroom-style presentation theme

warm-modern	minimal	tech
Warmer editorial look	Monochrome and restrained	Dark technical / product style

Quality Pipeline

prompt + sources -> planning -> schema validation -> PPTX render -> visual QA -> editable .pptx

Quality controls already in the repository include:

• JSON schema validation before rendering
• Chart repair and fallback when chart data is invalid
• Visual QA heuristics for overlap, edge crowding, and empty-slide detection
• Theme-aware text/background contrast handling across light and dark themes

What It Does

Entry Points

MCP Demo

Below is a terminal-style demo generated from a real MCP stdio round-trip against the current mcp_server.py implementation.

What it demonstrates:

• MCP server initialization over stdio
• Tool discovery (create_deck, revise_deck)
• A real create_deck tool call with mock: true
• Returned artifact paths for generated deck JSON and PPTX

MCP Configuration

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "auto-ppt": {
      "command": "python",
      "args": ["/absolute/path/to/auto-ppt-engine/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-engine/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-engine python mcp_server.py  # MCP stdio

Testing

python -m pytest tests/ -v   # 408 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

flowchart LR
  MCP["MCP\ncreate_deck / revise_deck"]
  CLI["CLI\ngenerate / revise"]
  HTTP["HTTP skill server\nPOST /skill"]
  JSON["JSON skill\nrequest file"]
  Sources["Prompt + sources"]
  Smart["Python smart layer\nplanning · revision · source loading\nschema validation"]
  Deck["Validated deck JSON"]
  JS["JS renderer\npptxgenjs + chart fallback"]
  PY["Python template renderer\npython-pptx"]
  QA["Visual QA / audit"]
  PPTX["Editable .pptx"]

  MCP --> Smart
  CLI --> Smart
  HTTP --> Smart
  JSON --> Smart
  Sources --> Smart
  Smart --> Deck
  Deck --> JS
  Deck --> PY
  JS --> QA
  PY --> QA
  QA --> PPTX
  PPTX -. revise loop .-> Smart

Default output goes through the JS renderer; the Python renderer is used for template-oriented rendering paths.

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

Documentation

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

Acknowledgments

Built with the assistance of Claude (Anthropic) and GitHub Copilot.

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE for the full text.