What it does

PPTAgent (now branded DeepPresenter) generates PowerPoint presentations from text prompts, attached documents, or raw data. It runs as a CLI tool, a local web UI, or a Docker Compose stack, and outputs real PPTX files rather than screenshots or markdown. The system can ingest PDFs, Excel files, and web search results, then structure them into slides with generated layouts and images.

The interesting bit

The project ships its own 9B parameter model, DeepPresenter, fine-tuned specifically for this task — the authors claim it “significantly outperforms existing open-source models” in their experiments. The runtime uses a sandboxed Docker container with 20+ tools for isolated execution, and recently added MCP server support so you can plug it into broader agent workflows.

Key highlights

• Fine-tuned 9B model (GGUF and full weights) available on HuggingFace and ModelScope
• Three deployment modes: CLI via uvx, source build, or Docker Compose with separate host/sandbox images
• Optional integrations: Tavily for search, MinerU for PDF parsing, text-to-image models for visuals
• Offline mode available if you run MinerU locally and disable network-dependent tools
• Papers accepted at EMNLP 2025 (PPTAgent) and ACL 2026 (DeepPresenter)

Caveats

• Windows is explicitly not supported; WSL required
• macOS CLI auto-installs a long list of dependencies (Homebrew, Node, Docker, Playwright, llama.cpp); Linux users must prepare the environment manually
• The README strongly recommends their fine-tuned model for best results, implying generic LLMs may struggle

Verdict

Worth a look if you regularly turn reports into decks and want an open, hackable pipeline. Skip it if you need a polished SaaS experience or if “install Docker, Node, Playwright, and a 9B model” sounds like too much Tuesday.