An AI agent that builds your deck and checks its own work

What it does

PPTAgent turns a prompt, document, or spreadsheet into a complete PowerPoint deck using an agentic loop that researches, designs layouts, and generates visual assets. The system runs inside a sandboxed environment with over 20 tools and exports finished work as PPTX files. A reflective critique step lets the agent revise its own slides rather than quitting after the first draft.

The interesting bit

Instead of relying on generic LLMs, the team ships a dedicated 9B parameter model called DeepPresenter—available quantized or full weights—and claims it significantly outperforms open-source alternatives for this specific task. The whole pipeline can also run fully offline by swapping web search and cloud parsing for local tools, which is a rare escape hatch in the current agent-tooling landscape.

Key highlights

• Reflective generation with self-critique and revision before final output
• DeepPresenter-9B, a fine-tuned model specifically for presentation tasks
• Fully offline mode via local sandbox and document parsing
• MCP server support for plugging into broader agent ecosystems
• Docker-isolated tool execution and context management to prevent overflow

Caveats

• Windows is not supported; WSL is required
• The authors strongly recommend their own fine-tuned model, so behavior with off-the-shelf LLMs is unclear
• The dependency stack spans Docker, Playwright, Node.js, and optional local inference engines

Verdict

A solid bet for developers who need to automate document-to-presentation workflows on their own infrastructure. Look elsewhere if you want a lightweight, cross-platform, or browser-only solution.