Cortex Slides is a standalone application (currently in beta) that turns the domain knowledge inside your Cortex instance into presentation decks — pitch, training, or report, 6 to 30 slides — delivered as a .pptx plus a PDF. Like Cortex Chat it runs as its own deployment and connects to your instance purely through the REST API with a scoped key (a standalone app, as opposed to Apps hosted inside an instance). It is the third sibling of Cortex Trainings and Cortex Videogen, built on the same architecture: a resumable file-backed pipeline, human gates before every spend, per-artefact caching, and two-channel research grounding.
Full-image slides
Slides are full-image: gpt-image-2 (via Venice.ai) paints each entire slide — typography included — from a prompt assembled out of four owned parts: the layout's composition brief, the slide's exact approved text quoted verbatim, its artwork concept, and a deck-wide design system. The result is a maximally designed page on every slide, in a class generative layout engines and template decks don't reach.
Every slide after the first approved sample is painted against that sample as a style anchor — the approved page travels along as a reference image, so palette, typography, panel language and texture stay identical across the deck instead of drifting page to page.
The trade is explicit: the text on the slides is painted pixels, not editable PowerPoint text. Painted text can come out wrong, so the pipeline is built around it — every word is approved before generation, QA tells you to proofread instead of pretending to verify pixels, and any single slide regenerates for the price of one image.
How it works
A production run has six steps and three human gates — content, look, and money:
- Research (free) — two-channel grounding against your instance: agentic Deep Research writes the argument, verbatim evidence excerpts from hybrid search preserve exact names and numbers through summarization, and the graph's topic communities are appended as outline fuel. Sectioned decks get one research pass per section.
- Outline (content gate) — the whole deck is drafted across 8 layouts (title, section, bullets, statement, image-full, two-column, quote, data), each slide with a takeaway title, its text, and an artwork concept. You review it next to the research it is grounded in — approve (optionally with per-slide hand edits, applied verbatim), regenerate with feedback, or re-run the research with steering. All free, unlimited rounds. This gate matters more than in any sibling app: every word approved here is painted into pixels.
- Style (look gate) — the app proposes three of its ~12 design-system presets with rationale, then paints one representative content slide in full as a sample — one paid image, cached per preset so flipping between styles never re-pays. What you approve is exactly a finished page, not a swatch, and it becomes the style anchor every later slide is generated against. Approved styles can be saved to a cross-project style library.
- Visuals (cost gate) — every missing slide is quoted against Venice.ai's live pricing and the run pauses with the total; nothing is generated until you confirm. The approved sample already counts as done. Optional subject reference images ride along on the slides that feature them, binding a product, mascot or character's identity while each page stages it in its own pose and camera angle.
- Assembly — pure packaging: the slide images become the
.pptx(one full-bleed image per slide) and the PDF. No generation happens here. - QA — structural probes (slide count, images embedded, PDF pages), ending with what it deliberately does not claim: painted text cannot be machine-verified — proofread the deck, and regenerate individual slides to fix errors.
Everything is cached per artefact: a crashed or resumed run never pays for the same slide twice. After delivery, a text or artwork fix on one slide quotes and repaints just that slide, always as a fresh full render from your corrected words rather than a touch-up of the existing image.
Setup
Cortex Slides is a Node.js application (Node 22+). You need a Venice.ai API key for generation and a read-scoped API key for your Cortex instance:
Code
| Variable | What it is |
|---|---|
VENICE_API_KEY | Venice.ai key — pays for research/outline LLM calls and slide images |
CORTEX_BASE_URL | Your instance, e.g. https://cortex.example.com |
CORTEX_API_KEY | A read-scoped API key for it |
Optional: ACCENT_COLOR (hex) and BRAND_NAME brand the presets that use them; APP_LANG switches the UI (en/de) — the deck's content language is chosen per project; IMAGE_RESOLUTION (1K/2K) and IMAGE_QUALITY pick the image tier. npm run smoke renders a synthetic all-layouts deck through the real packaging path for $0 — the instance needs ENABLE_AGENTIC_RAG and ENABLE_AGENT_RESEARCH for the research step.
Create a project in the form: topic, audience, purpose (pitch/training/report), target language, slide count, text density, optional call-to-action (carried verbatim onto the final slide), optional sections for long trainings, optional subject/style reference images, optional collection scope.
Costs
Slide images are the only real spend; research and outline are cents of LLM tokens. Reference: ~$0.50 per slide on gpt-image-2 at 2K/high, so a 12-slide deck lands around $6 — the quote gate shows the exact number before anything is generated, the style sample costs one image per round, and fixing one slide later costs one image.
Learn more
- Repository & full docs — pipeline internals, Venice provider notes, the full-image design rationale
- Cortex Trainings — the sibling app for interactive training courses
- Cortex Videogen — the sibling app for marketing videos
- Apps — the in-instance app system this is not (Slides runs standalone)