A new open-source project called Codex Slides turns a plain prompt, a folder of files, or an entire code repository into a finished presentation deck, and it runs entirely inside OpenAI's Codex coding agent. Released by the team at nexu-io under an MIT license, the project had already gathered close to 400 GitHub stars within two weeks of going public, and it landed on Hacker News on July 24, 2026. The pitch is specific: 10 or more high-quality, image-native slides in roughly four to five minutes, with every step of the process visible in the browser as it happens.
What makes Codex Slides worth a closer look is not that it generates slides, plenty of tools do that, but where it sits. Instead of being a standalone web app with its own model and its own account, it is a plugin that lives inside the coding agent you may already be using to write software. That changes the workflow in ways that matter for anyone who builds things and then has to explain them.
What Codex Slides Actually Does
Codex Slides is what its authors call an "image-native" deck studio. Rather than filling a text template with bullet points, it treats each slide as a full visual canvas and composes the whole page, layout, imagery, typography, and data together. It ships with 118 pre-built visual directions: 45 curated deck templates across 15 professional categories, plus 73 community-contributed styles spanning reports, infographics, diagrams, data dashboards, posters, product showcases, brand systems, architecture renders, and editorial layouts.
Because it operates as a plugin inside OpenAI Codex, it can read the same files, repositories, and context that the agent already has access to. Point it at a GitHub repo and it can pull the README, the architecture, and the commit history into a technical deck. Give it a research folder and it can outline a findings presentation. The output is not a rough draft you export and fix elsewhere: Codex Slides produces production-ready PPTX and PDF files with multiple quality levels and aspect ratios, plus a dual-window presentation mode with speaker notes.
How the Repo-to-Deck Workflow Works
The distinguishing feature is that you watch the entire chain run live in a persistent browser interface, rather than waiting for a finished file to drop. The stages run in order, and you can intervene at any point:
- Research. Codex Slides gathers source material from your prompt, your attached Design Files, or the repository the agent is pointed at.
- Outline. It proposes a slide-by-slide structure you can edit before any rendering happens.
- Style. You pick one of the 118 visual directions, or let it match your saved Brand and Design System.
- Render. In Fast mode, every page renders in parallel rather than one at a time, which is how the four-to-five-minute figure for 10-plus slides is reached.
- Edit. A mark-based annotation system lets you point at a specific element and describe the change, instead of regenerating the whole slide.
- Present and export. Run the dual-window presenter with speaker notes, or export to PPTX and PDF.
The plugin uses the Model Context Protocol to talk to Codex, and projects are durable and stored locally, so a deck you started yesterday is still there today. That local-first, browser-first design means there are no separate API keys to manage: authentication rides on your existing Codex login.
Codex Slides vs Other AI Deck Tools
The AI presentation category is crowded, but most tools are hosted web apps that own your content and charge per seat. Codex Slides is aimed at a different user: the developer or technical creator who already works inside a coding agent and wants decks generated from real project context. Here is how the trade-offs compare.
| Factor | Codex Slides | Hosted AI deck tools |
|---|---|---|
| License | Open source (MIT) | Proprietary, subscription |
| Where it runs | Inside Codex, local/browser-first | Vendor cloud |
| Source context | Reads your repos and files directly | Manual prompt or upload |
| API keys | None; uses Codex login | Account required |
| Export | Editable PPTX and PDF | Varies; often PDF or lock-in |
| Editing | Mark-based, per-element annotation | Regenerate or manual |
For comparison, a hosted tool like Beautiful.ai gives you polished templates and team features but keeps your decks on its servers, and a tool such as Brightdeck focuses on producing editable PPTX from a prompt. Codex Slides trades some of that hosted polish for control, portability, and direct access to the code and files you are already working with.
What This Enables
The practical payoff is speed on the least fun part of building: explaining what you built. If you have shipped a feature, the deck for the demo, the investor update, or the internal review can be generated from the repo itself, so the slides reflect what the code actually does rather than a hand-written summary that drifts out of date. Because projects are durable and locally stored, you can keep a living deck for a project and re-render it as the code changes.
For teams, the Brand and Design System persistence means one saved style can be reused across every deck, so a solo developer or a small studio can ship on-brand presentations without a designer in the loop. And because the output is standard PPTX and PDF, the deck drops into whatever review or sharing process a company already uses. This is the same "bring the tool to where the work is" pattern behind other Codex-native projects, like a local-first Figma clone that also builds on Codex and MCP.
Setup and Requirements
Codex Slides is built with Next.js 14 and TypeScript, and it operates as a Codex plugin with MCP support and a portable command-line interface. To run it you need Node.js 20 or newer, a Codex install with plugin support, and a ChatGPT login handled through codex login. There is no separate model subscription or image-generation API key to configure, which lowers the barrier for anyone already inside the Codex ecosystem.
Because it is open source, you can read exactly how the rendering pipeline works, fork it, or contribute one of the community visual styles. The project sits alongside the growing set of tools built on the Codex CLI, and its 29 forks in two weeks suggest active early interest from developers who want to extend it.
Limitations and What to Watch
The obvious constraint is that Codex Slides only makes sense if you use Codex. It is not a general web app for non-technical users, and the ChatGPT-authentication requirement ties it to that account. As a two-week-old open-source project it will have rough edges, and the visual quality of image-native slides depends heavily on the underlying generation, which can vary. It is also worth weighing against other Codex-ecosystem routing tools, such as OpenCodex, if you want more control over which model does the work. For creators who live in a terminal and a repo, though, generating a real deck without leaving that environment is a genuinely new option.
Frequently Asked Questions
Is Codex Slides free?
Yes. It is open source under the MIT license, so you can use, modify, and self-host it at no cost. You do need a working Codex install and a ChatGPT login, which have their own terms.
Do I need API keys to use it?
No separate API keys are required. Authentication rides on your existing Codex login via codex login, and the project is designed to be browser-first with locally stored projects.
What file formats can it export?
Codex Slides exports production-ready PPTX and PDF files, with multiple quality levels and aspect ratios, so the decks are editable in PowerPoint, Keynote, or Google Slides after export.
How fast is it?
The project advertises 10 or more slides in roughly four to five minutes in Fast mode, which renders every page in parallel rather than sequentially.
Can it build a deck from a code repository?
Yes. Because it runs as a plugin inside Codex, it can read the repositories and files the agent has access to and turn them into a technical presentation, which is the main reason it is aimed at developers.
What do I need to run it?
Node.js 20 or newer, a Codex install with plugin support, and a ChatGPT login. It is built with Next.js 14 and TypeScript and supports the Model Context Protocol.
