AI-driven PowerPoint generation, audited slide-by-slide, ready for Claude Code and any MCP-compatible client.
Megadeck turns a natural-language prompt into a styled .pptx file. It exposes the same core both as a command-line tool and as a Model Context Protocol (MCP) server, so Claude Code, Claude Desktop, and any MCP-aware client can call it directly.
It is opinionated about three things:
- Constrained output — the LLM never produces freeform text; it produces a typed Pydantic schema that the renderer turns into pptx.
- Visual auditing — every slide is rendered to PNG and checked for overflow, alignment, and sizing issues before the deck is saved.
- Theming first — slide content lives in JSON; visual style lives in themes. Switch themes without rewriting content.
The fastest path is uvx (zero install, runs the latest from PyPI):
# CLI
uvx megadeck generate "30-slide masterclass on entrepreneurship and AI agents" \
--output deck.pptx
# MCP server (stdio)
uvx megadeck-mcpOr install with pip:
pip install megadeck
megadeck --helpclaude mcp add megadeck \
--command "uvx megadeck-mcp" \
--env "ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY"Then in any Claude Code session:
"Read
notes.mdand generate a 20-slide pitch deck. Use the editorial theme."
Claude calls Megadeck's MCP tools, runs the critic loop, and saves the .pptx.
Megadeck speaks the standard Model Context Protocol over stdio, so any MCP-aware client (Claude Desktop, Continue, Zed, etc.) can launch it. Drop the following into the client's MCP config:
{
"mcpServers": {
"megadeck": {
"command": "uvx",
"args": ["megadeck-mcp"],
"env": { "ANTHROPIC_API_KEY": "..." }
}
}
}Restart the client; its agent now has access to Megadeck.
| Template | When to use |
|---|---|
hero_statement |
One-line bold idea, e.g. "Hard is good." |
numbered_list |
Up to six points with big outlined numbers |
three_card |
Three side-by-side concepts |
two_column |
Comparison / before-after |
section_divider |
Part-break with eyebrow + giant title |
More templates land in Phase 2 — see ROADMAP.md.
| Theme | Mood |
|---|---|
default |
Sky-blue, minimalistic, lots of whitespace |
dark |
Slate-900 background, white type |
editorial |
Serif headings, neutral palette |
megadeck generate "<prompt>" # full pipeline
megadeck audit deck.pptx # render + critique an existing deck
megadeck themes # list available themes
megadeck templates # list available slide templates
megadeck preview deck.pptx --out png/ # render every slide to PNGWhen loaded via megadeck-mcp, the server exposes these tools:
generate_deck(prompt, theme, n_slides, output_path)audit_deck(path)— returns per-slide overflow/alignment reportlist_templates()list_themes()
Picks up whichever environment variable is set, in this order:
ANTHROPIC_API_KEY→ Claude Sonnet 4 (default)OPENAI_API_KEY→ GPT-4oGOOGLE_API_KEY→ Gemini 2.5 Pro (withpip install megadeck[google])
You can force a provider with --model anthropic|openai|google.
prompt
│
▼
┌──────────────────────────────────────────────────┐
│ LLM + instructor → validated Pydantic Deck │
│ (constrained output — no malformed JSON ever) │
└──────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ Renderer (theme-aware) → python-pptx → .pptx │
└──────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ Preview (libreoffice → PDF → PNG per slide) │
└──────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ Critic (LLM audits PNG for overflow/alignment) │
│ → flagged slides regenerate until they pass │
└──────────────────────────────────────────────────┘
│
▼
.pptx (clean, audited, ready to present)
See ROADMAP.md. Phase 1 ships the foundation, the MCP/CLI surface, and 5 production templates. Phase 2 adds 15+ more templates, animation primitives, and additional themes. Phase 3 adds a web preview UI and Smithery registry submission.
python-pptx— pptx file authoringpydantic— schema validationinstructor— typed LLM outputfastmcp— MCP server frameworktyper+rich— CLI surface- LibreOffice (headless) +
pdf2image— slide preview - Architecture inspired by AI-assisted code-generation tools and constrained-output prompting techniques.
See CONTRIBUTORS.md.
MIT — see LICENSE.