If you have ever asked an AI to "make slides" and received a deck of flattened images or a template with text dropped in, PPT Master is aimed squarely at you. The open-source project, built by Hugo He and trending on GitHub this week, generates real PowerPoint files: native shapes, data-backed charts and tables, transitions, animations and audio narration from speaker notes, all editable in PowerPoint afterwards. It is not an app but a "skill", a workflow that an agent such as Claude Code runs on your machine.
This guide covers what it does, how to install it, what it costs, where it falls short, and how it compares to other AI slide tools we have covered.
Quick answers
| Question | Answer |
|---|---|
| What is it? | An open-source agent skill that exports native, editable .pptx |
| Who made it? | Hugo He, a finance professional (CPA) who wanted editable AI slides |
| Cost? | Free; you pay only for model usage |
| Needs? | Python 3.10+ and an agent-capable AI tool |
| Recommended model? | Kimi K3 or Claude, with gpt-image-2 or Nano Banana 2.1 for images |
| Data privacy? | Runs locally apart from calls to the AI model |
| Catch? | You still polish the output; cheaper models need more cleanup |
What makes it different
The README argues "editable is already table stakes", and that what separates tools is how much of PowerPoint's object model you actually get. PPT Master emphasizes native shapes and connectors with working adjustment handles, data-backed charts and tables on demand, and the full text, picture, fill and effect model. Through its template route it can produce real slide masters and layouts (the p:sldMaster and p:sldLayout inheritance PowerPoint uses). Formulas compile to editable OMML equations for PowerPoint 2010 and later.
It also tries to reason the argument into shape before designing: the agent analyzes the source, structures the narrative, confirms a design spec with you, then generates slides as SVG and exports them to DrawingML. The project documents a PowerPoint-to-SVG mapping guide as an honest, feature-by-feature record of what is supported; SmartArt is listed as a deliberate omission.
Beyond generating new decks, the README lists other routes:
- Distilling reusable brand, style, layout and deck templates from your references.
- Filling an existing .pptx with new content while preserving the design (unchanged pages are kept byte-for-byte).
- Adding native transitions, animations and narration to a finished deck.
- A quick mode that skips the confirmation round trip.
Setup in four steps
The README's quick start is short.
- Install Python 3.10 or newer. Windows users get a dedicated guide because of PATH and execution policy quirks.
- Pick an agent. The author's pick is Claude Code, but Cursor, Codex CLI, Gemini CLI, Copilot, Cline and others work.
- Get the project:
git clone https://github.com/hugohe3/ppt-master.git, thenpip install -r requirements.txt. Alternatively install it as a skill:npx skills add hugohe3/ppt-master, or inside Claude Code run/plugin marketplace add hugohe3/ppt-masterand/plugin install ppt-master@ppt-master. - Open the folder in your agent and chat: "Please create a PPT from projects/q3-report/sources/report.pdf".
By default the agent first proposes a design spec (template, format such as 16:9, page count) before generating. Say so explicitly and it will skip confirmation. The deck lands in an exports folder as a timestamped .pptx, with page previews alongside. A flag produces data-backed PowerPoint chart and table objects with Edit Data as a separate file.
Our primer on what agent skills are explains why a skill like this works across tools, and how to turn agent skills into loops shows the next step if you want decks generated on a schedule.
What does it cost?
The software is free. Your real cost is the model. The README recommends Kimi K3 or Claude with roughly a one million token context, plus an image model. Kimi is also a project sponsor, and the README lists several API relay sponsors, so read recommendations with that in mind. If you want to run Kimi cheaply, see our coverage of Kimi K3 open weights and the Kimi K3 API guide.
For images, the tool can use the agent's built-in image tool, an image-generation script with a provider key, or a web image search that works with zero config using Openverse and Wikimedia Commons, and better with free Pexels and Pixabay keys. It handles license attribution, adding a credit when the chosen image requires one. For image model pricing context see Google's Nano Banana 2.1 price cut and GPT Image 2 transparent backgrounds.
Honest limitations
The author is blunt in a box titled "This is a tool, not a wishing well": "harness plus model equals agent", so the model sets the ceiling. Expect to polish the deck. If results disappoint, upgrade the model first, then check your usage against the docs.
Other cautions from reading the README:
- It needs an agent environment; there is no hosted web app or one-click GUI. Non-technical users must install Python and an AI tool.
- Legacy formats like .doc, .rtf or .tex need pandoc.
- Image quality depends on provider keys; zero-config search can return uneven images.
- SmartArt is not supported by design.
- Sponsor relationships (API relay services with discount codes) appear prominently; evaluate them independently.
- The repo has more than two thousand commits and moves quickly, so pin a version if you automate it.
What people will ask next
Is the output really different from a template fill? The README says it is not "a filled-in template". The agent writes each slide as SVG and converts it to PowerPoint DrawingML, so shapes, connectors and text boxes are individual objects. The README shows three example decks, each a single pass with no manual polish: a pixel-art atlas of Chinese breakfast with sprite sheets and 8-bit sound cues, a technical blueprint walking through the Transformer paper with native formulas and animations, and a brand-template deck built on a company's own template with native-chart export. Downloading one and opening it in PowerPoint is, in the author's words, the fastest way to see what it can do.
How do I update it? Git clone installs run a bundled update script that pulls the latest version and syncs Python dependencies when requirements change. ZIP installs cannot pull, so you download a fresh copy, copy over your .env and projects folder, and reinstall requirements. A skill-only package of roughly 56 MB, without the bundled example decks, is offered on the Releases page if the full download fails.
Where do my API keys live? In a .env file. The tool reads the process environment first, then a .env in the working directory, the skill directory, the clone root, or a persistent user config at ~/.ppt-master/.env. Keep it out of version control like any other credential file.
What if the agent loses context mid-run? The README says to ask it to re-read the skill's SKILL.md file, and to check the FAQ, which the author says is continuously updated from real user reports.
Can it use my company template? Yes, through the template route: you provide a reference deck and the tool distills brand, style and layout, or you ask it to fill an existing deck. The brand-template example is the quickest way to judge how faithful that is.
Why the "skill" form matters
A shelf of skill cards with one pulled out, illustrating PPT Master as an installable agent skill
Most AI slide products are websites: you upload, they render, you export. PPT Master inverts that. Because it is a markdown-described workflow plus Python scripts, any agent harness that can read files and run commands can drive it, and you can read exactly what it will do before it does it. That makes it auditable and forkable, and it means a team can wrap it into an internal pipeline, for example generating a weekly report deck from a data export. The cost is that you carry the setup burden and the quality depends on the model you choose.
It also reflects a broader pattern we track: useful AI tools shipping as skills and plugins rather than apps, distributed through marketplaces and installed in an agent you already pay for. If you are new to that pattern, start with our guide to agent skills and the Karpathy-style guideline file linked below.
Practical tips before your first deck
- Start with a short source document, a few pages, to see the design-spec step and the visual baseline without burning many tokens.
- Use a large-context model; long reports plus images eat the window quickly.
- Decide up front whether you need native PowerPoint chart objects with Edit Data, since they come out as a separate file.
- Open the result in PowerPoint, not just a previewer, before sending it to anyone; this is where native shape problems show up.
- Keep sources and the generated deck together under the projects folder so the project stays reproducible.
How it compares
PPT Master's closest analogue in our coverage is Kimi Slides, a hosted research-to-PPTX flow that also stays editable. The trade-off is hosted convenience versus local control: Kimi Slides is a product, PPT Master is a workflow you run and can modify. Microsoft's own route is Copilot inside the apps, which we covered when Grok joined Copilot in Word, Excel and PowerPoint.
| Option | Where it runs | Editable native output | Cost |
|---|---|---|---|
| PPT Master | Your machine, any agent | Yes, native shapes, charts, masters | Free plus model usage |
| Kimi Slides | Hosted | Yes | Kimi plan |
| Copilot in PowerPoint | Inside Microsoft 365 | Yes | Microsoft subscription |
Prompts to try
- "Create an 8 to 10 slide deck from projects/q3-report/sources/report.pdf, 16:9, with native charts."
- "Quickly generate a 5-page deck from this text, no need to confirm with me."
- "Fill this existing deck with the new content and keep the design."
- "Add transitions and narration from the speaker notes."
Should you adopt it?
If you regularly turn reports into decks, already use Claude Code or a similar agent, and care that the output stays editable, PPT Master is worth an afternoon. If you want a click-and-go web tool, a hosted product is easier. As always with agent skills, review what a skill can run on your machine before installing; our guide to the Karpathy-inspired Claude Code skills shows how much behavior a few markdown files can steer.
Details reflect the project README as of October 10, 2026 and may change.