PPTAgent Review: Open Source AI Tool Automatically Converts Documents to Professional PPT

Q: How to deploy PPTAgent on a company intranet?

PPTAgent supports fully offline mode. Pull Docker images, download DeepPresenter-9B GGUF model locally, deploy MinerU local PDF parsing, and set offline_mode: true in config.yaml. All processing happens locally.

Q: How does PPTAgent differ from Manus/ChatGPT PPT features?

PPTAgent's core differences: 1) Deep document understanding — directly parses PDF/Excel; 2) Reflection mechanism — self-checks and corrects after generation; 3) Open source and controllable — data stays local, models are replaceable.

PPTAgent Review: Open Source AI Tool Automatically Converts Documents to Professional PPT

2026-09-03 PPTAgent AI PPT生成文档转PPT 开源PPT工具 DeepPresenter LLM

Three Pain Points of PPT Making That Every Knowledge Worker Knows

Monday morning department meetings, end-of-quarter OKR reviews, client presentations before project delivery — these scenarios share a common nightmare: making PPTs.

According to Microsoft's 2026 Work Trend Index report, knowledge workers spend an average of 3.2 hours per week on presentation creation, with 60% of that time spent on layout adjustments rather than content creation. The harsh reality:

Content extraction is time-consuming: Distilling 15 pages of精华 from a 50-page technical document requires repeated reading, summarizing, and restructuring
Layout design has a high barrier: Color schemes, alignment, image-text ratios — non-designers struggle to create "presentable" slides
Repetitive labor is severe: Every report follows the same dead loop: "find template → fill content → adjust formatting"

The market is not short of PPT tools, but truly open source solutions that can understand document content and automatically generate structured presentations only emerged with PPTAgent.

What Is PPTAgent: More Than Just "Document to PPT"

PPTAgent is an open source AI presentation generation framework from the Institute of Computing Technology, Chinese Academy of Sciences (ICIP-CAS), with 4,900+ Stars on GitHub. It's not a simple template-filling tool but a multi-agent collaborative system (Agentic Framework) that can:

Deeply understand documents: Parse PDF, Word, Excel formats, automatically extracting core arguments and data
Reflective generation: Iterate through Research → Design → Reflect loops to continuously optimize output quality
Image-text mixing: Automatically call text-to-image models for illustrations, or extract charts from documents
Multi-format export: Support for native PPTX format and HTML slides

Document Input → Research Agent (Content Research) → Design Agent (Visual Design) → Reflection Optimization → PPTX Output

In August 2025, the PPTAgent paper was accepted to EMNLP 2025; in April 2026, its fine-tuned model DeepPresenter was accepted to ACL 2026 — a dual recognition from the academic community.

Core Feature Matrix: What PPTAgent Can Do

Feature Module	Description	Technical Implementation
Document Understanding	Parse PDF/Word/Excel/Markdown	MinerU + MarkItDown
Content Research	Auto-search supplementary info	Tavily Web Search + arXiv
Smart Layout	Choose best layout by content type	Design Agent + HTML Rendering
Image Generation	Auto-generate illustrations, charts, flowcharts	T2I Models + Mermaid + Plotly
Reflective Optimization	Detect and fix content/layout issues	Reflection Loop (up to 10 rounds)
MCP Server	Be called by other Agents as a tool	FastMCP Server
Web UI	Browser-based visual interface	Gradio (port 7861)
CLI Mode	Quick command-line generation	Typer CLI

Three Operation Modes

1. CLI Mode (Recommended for Personal Use)

# Install
uv tool install pptagent

# Interactive setup wizard
pptagent onboard

# Generate PPT with one sentence
pptagent generate "Q4 Quarterly Report" -o report.pptx

# Generate with attachments
pptagent generate "Technical Review" \
  -f architecture.pdf \
  -f benchmark.xlsx \
  -p "10-15" \
  -o tech-review.pptx

2. Docker Compose (Server Deployment)

docker pull forceless/deeppresenter-sandbox
docker pull forceless/deeppresenter-host
docker tag forceless/deeppresenter-sandbox deeppresenter-sandbox
docker tag forceless/deeppresenter-host deeppresenter-host

docker compose up -d
# Visit http://localhost:7861

3. Source Build (Developer Mode)

git clone https://github.com/icip-cas/PPTAgent.git
cd PPTAgent
uv pip install -e .
playwright install-deps && playwright install chromium
npm install --prefix deeppresenter/html2pptx
python webui.py

Technical Architecture Deep Dive

PPTAgent's core design philosophy is "Reflective Generation" — not one-shot output, but approaching optimal results through multi-round iteration.

Multi-Agent Collaboration Flow

┌─────────────────────────────────────────────────┐
│                  AgentLoop.run()                 │
├─────────────────────────────────────────────────┤
│                                                 │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐  │
│  │ Research │───▶│  Design  │───▶│ Reflect  │  │
│  │  Agent   │    │  Agent   │    │  Loop    │  │
│  └──────────┘    └──────────┘    └──────────┘  │
│       │               │               │        │
│       ▼               ▼               ▼        │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐  │
│  │  Web     │    │  HTML    │    │  Quality │  │
│  │  Search  │    │  Render  │    │  Assess  │  │
│  │  arXiv   │    │  PPTX    │    │  Fix     │  │
│  └──────────┘    └──────────┘    └──────────┘  │
│                                                 │
└─────────────────────────────────────────────────┘

Research Agent handles the content layer: parsing input documents, optionally searching the web for supplementary knowledge, and generating a structured "manuscript". It supports context management — when documents are too long, it automatically folds early content (context_folding: true) to avoid exceeding the model's context window.

Design Agent handles the visual layer: selecting appropriate layouts based on content type (technical report/product introduction/teaching courseware), generating HTML slide code, then rendering to images via Playwright or converting directly to PPTX.

Reflection Loop is key to quality assurance: the system evaluates completeness, logic, and visual consistency of generated content, automatically triggering corrections when issues are found, iterating up to 10 rounds (MAX_SUBAGENT_TURNS).

Sandbox Isolation

PPTAgent uses Docker sandboxes for code generation tasks (chart drawing, data processing), preventing malicious code from affecting the host environment:

# Sandbox timeout control
MCP_CONNECT_TIMEOUT: 120  # Connection timeout (seconds)
MCP_CALL_TIMEOUT: 1800    # Call timeout (seconds)

# Tool call limits
MAX_TOOLCALL_PER_TURN: 7  # Max 7 tool calls per turn
TOOL_CUTOFF_LEN: 4096     # Tool output truncation threshold (chars)

Flexible Model Configuration

PPTAgent doesn't bind to any specific LLM provider, supporting 100+ models through OpenRouter or LiteLLM:

# config.yaml example
research_agent:
  base_url: "https://openrouter.ai/api/v1"
  model: "anthropic/claude-sonnet-4.5"
  api_key: "your_key"

design_agent:
  base_url: "https://openrouter.ai/api/v1"
  model: "google/gemini-3-pro-preview"
  api_key: "your_key"

# Also supports LiteLLM routing
# research_agent:
#   provider: "litellm"
#   model: "anthropic/claude-sonnet-4-6"
#   api_key: "your_anthropic_key"

The team strongly recommends using the fine-tuned model DeepPresenter-9B (downloadable from HuggingFace), which significantly outperforms general-purpose open source models on PPT generation tasks.

PPTAgent vs Gamma vs Beautiful.ai vs ChatBA: Horizontal Comparison

There are plenty of AI PPT tools on the market, but their positioning varies greatly. Here's a comparison of mainstream solutions in 2026:

Dimension	PPTAgent	Gamma	Beautiful.ai	ChatBA
Open Source	✅ MIT License	❌ Closed SaaS	❌ Closed SaaS	❌ Closed SaaS
Deployment	Local/Docker/Source	Online SaaS	Online SaaS	Online SaaS
Document Understanding	Deep (PDF/Word/Excel)	Medium (text paste)	Weak (plain text)	Weak (plain text)
Image-Text Mixing	✅ Auto images+charts	✅ Template library	✅ Smart layout	⚠️ Basic images
Reflective Optimization	✅ Multi-round iteration	❌ One-shot	❌ One-shot	❌ One-shot
Custom Models	✅ 100+ LLMs	❌ Fixed model	❌ Fixed model	❌ Fixed model
MCP Integration	✅ Agent tool	❌	❌	❌
Data Privacy	✅ Fully local	❌ Data uploaded	❌ Data uploaded	❌ Data uploaded
Price	Free (bring your API key)	From $10/mo	From $12/mo	Free tier+paid
Target Users	Developers/Tech teams	Business/Marketing	Designers	Quick prototyping

Key Differences:

PPTAgent is the only open source option — fully local data, suitable for enterprises with privacy requirements
PPTAgent's reflection mechanism is a unique advantage — other tools are "one-shot", PPTAgent self-checks and corrects
Gamma/Beautiful.ai win on ease of use — open a webpage and start, no environment setup needed
PPTAgent has the strongest document understanding — drop in a 50-page PDF and it extracts key points automatically

Hands-On: Full Workflow from Technical Document to Presentation PPT

Let's walk through a complete PPTAgent workflow using a real technical whitepaper.

Step 1: Environment Setup

# Install uv (Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install PPTAgent
uv tool install pptagent

# First-time setup (interactive wizard, configure LLM API Key)
pptagent onboard

pptagent onboard guides you through: 1. Choosing an LLM provider (OpenRouter/OpenAI/Custom) 2. Entering your API Key 3. Optional configuration: Tavily (enhanced search), MinerU (enhanced PDF parsing), T2I model (enhanced illustrations)

Step 2: Generate PPT

# Basic usage: one sentence generation
pptagent generate "Explain the core principles of Transformer architecture" -o transformer.pptx

# Advanced usage: generate from documents
pptagent generate "Please create a technical report PPT based on the attached documents" \
  -f ~/docs/annual-report-2026.pdf \
  -f ~/data/metrics.xlsx \
  -p "12-18" \
  -o annual-review.pptx

Parameter explanation: - -f: Attachment path, supports PDF/Word/Excel/Markdown - -p: Target page range (e.g., "12-18" means 12 to 18 pages) - -o: Output file path

Step 3: Review and Adjust

The generated PPTX file can be opened directly in PowerPoint, WPS, or Google Slides for editing. PPTAgent's output characteristics:

Clear structure: Each page has a distinct theme with logical progression
Image-text matching: Key concepts include auto-generated diagrams or data charts
Consistent style: Colors and fonts are globally consistent

If adjustments are needed, you can modify the prompt and regenerate, or manually edit the PPTX file.

Advanced Usage: Custom Templates, Batch Processing, and API Integration

Custom Templates

PPTAgent supports importing custom PPTX templates to match corporate VI standards:

# Place template in workspace
cp company-template.pptx ~/.cache/deeppresenter/templates/

# Generate with template
pptagent generate "Product Introduction" \
  --template company-template.pptx \
  -o product-intro.pptx

Master slide layouts, color schemes, and font settings are preserved — PPTAgent only fills content areas.

Batch Processing

For scenarios requiring batch PPT generation (e.g., auto-generating presentations for each paper):

import subprocess
import glob

papers = glob.glob("papers/*.pdf")
for paper in papers:
    output = paper.replace(".pdf", ".pptx").replace("papers/", "presentations/")
    subprocess.run([
        "pptagent", "generate",
        "Create a 15-minute academic presentation based on this paper",
        "-f", paper,
        "-p", "15-20",
        "-o", output
    ])
    print(f"Generated: {output}")

MCP Integration: Let Other AI Agents Call PPTAgent

PPTAgent provides an MCP Server mode, callable by other Agents (like OpenClaw, Hermes):

// mcp.json configuration
{
  "mcpServers": {
    "pptagent": {
      "command": "pptagent-mcp",
      "args": [],
      "env": {
        "DEEPPRESENTER_WORKSPACE_BASE": "~/.cache/deeppresenter"
      }
    }
  }
}

Once configured, other Agents can call PPTAgent's presentation generation capability via the MCP protocol, enabling "conversational PPT creation".

Fully Offline Mode

For data-sensitive scenarios, PPTAgent supports fully offline operation:

# config.yaml
offline_mode: true

# Deploy local PDF parsing (MinerU)
# Deploy local LLM (e.g., DeepPresenter-9B GGUF)
# Configure local T2I model

In offline mode, all processing happens locally with no data sent to external services.

Limitations and Caveats

PPTAgent is not a silver bullet. Here are known limitations of the current version:

No native Windows support: Windows users need WSL (Windows Subsystem for Linux)
LLM API costs: Generating a 15-page PPT consumes roughly 50K-100K tokens, about $0.5-1.5 per run with GPT-4 level models
Slow first launch: Docker images need downloading (sandbox + host ~3GB), subsequent launches are fast
Occasional Chinese layout issues: Some Chinese fonts may have alignment issues during HTML rendering — fine-tune in PowerPoint afterward
Limited complex chart support: Highly customized data visualizations (e.g., interactive charts) still need manual handling
Optional services recommended: Tavily search and MinerU PDF parsing significantly improve quality but need extra API keys

Performance Benchmarks

Real-world benchmarks on M2 MacBook Pro (16GB RAM):

Document Size	Output Pages	Time	Token Usage
5-page PDF	8-page PPT	~2 min	~30K tokens
20-page Word	15-page PPT	~5 min	~65K tokens
50-page paper	20-page PPT	~8 min	~95K tokens

Final Verdict

PPTAgent is currently the most powerful open source AI PPT generation framework, period. Its core advantages:

Academic-grade methodology: Backed by EMNLP 2025 + ACL 2026 dual top-conference papers
Reflective generation: Not one-shot output — multi-round iterative optimization
Fully open source: MIT license, local data, freely customizable
Rich ecosystem: CLI/Web UI/MCP Server three modes, covering individual to team scenarios

Recommended for: - Developers/researchers who frequently create technical presentations - Enterprise users with data privacy requirements - Agent developers wanting to integrate PPT generation into AI workflows - Teams wanting to build on open source solutions

Not recommended for: - Casual users who make PPTs occasionally (Gamma is simpler) - Brand presentations needing极致 design (still requires professional designers) - Users without LLM API keys who don't want to configure

PPTAgent represents a new paradigm in AI PPT generation: not replacing humans in making PPTs, but understanding human needs and automatically generating high-quality drafts. You still need to review and adjust, but the repetitive "0 to 0.8" labor can finally be handed off to AI.

FAQ

1. Is PPTAgent free?

PPTAgent itself is completely free and open source (MIT license), but running it requires an LLM API. Using services like OpenRouter incurs API costs — generating a 15-page PPT costs about $0.5-1.5. If you deploy DeepPresenter-9B locally, you can achieve zero-cost operation (requires sufficient GPU VRAM).

2. What languages does PPTAgent support?

PPTAgent supports Chinese and English document parsing and PPT generation. For other languages (Japanese, Korean, etc.), content generation is basically usable but layout may need manual adjustment. The Web UI supports 21 languages.

3. How is the quality of generated PPTs? Do I need to manually edit?

PPTAgent generates PPTs with good structure and content quality, suitable for internal tech talks or first drafts. But for formal client presentations or brand showcases, we recommend fine-tuning in PowerPoint (adjusting fonts, replacing images, optimizing animations). The reflection mechanism significantly improves first-generation quality.

4. How to deploy PPTAgent on a company intranet?

PPTAgent supports fully offline mode. Steps: 1) Pull Docker images in a networked environment; 2) Download DeepPresenter-9B GGUF model locally; 3) Deploy MinerU local PDF parsing service; 4) Set offline_mode: true in config.yaml. All processing happens locally with no external data transmission.

5. How does PPTAgent differ from Manus/ChatGPT's PPT features?

Manus and ChatGPT's PPT features are essentially "text → PPT" — they understand your text description and fill templates. PPTAgent's core differences: 1) Deep document understanding — directly parses complex documents like PDF/Excel; 2) Reflection mechanism — self-checks and corrects after generation; 3) Open source and controllable — data stays local, models are replaceable.

PPTAgent in-depth review: an open-source AI-driven PPT generation framework supporting document understanding, intelligent layout, and image-text mixing. Compared with Gamma/Beautiful.ai, with complete installation guide and real-world case studies.