Documentation

autopresenter docs

One Markdown composition + one design system → on-brand artifacts. This is the working reference; the full specs live in the repo.

• What it is
• Install
• Composition
• Design system
• CLI
• Video
• Architecture

What it is

autopresenter is a local-first, agent-native studio. A short business brief generates a machine-readable DesignSystem (visual tokens + verbal identity); every artifact — slides, a narrated demo video, a GIF, and a landing page — inherits it, so output is on-brand by construction. Slides are one projection of a shared brand IR, not the product.

Install & run

git clone https://github.com/OpenHackersClub/autopresenter && cd autopresenter
node src/cli.ts render examples/autopresenter/deck.md --target slides -o dist/
node src/cli.ts present examples/autopresenter/deck.md   # live hot-reload preview

Requires Node ≥ 22.18 (native TypeScript). The slides projector has zero npm runtime dependencies.

The composition format

A composition is Markdown with ::: directives. Plain paragraphs are narration (→ presenter notes / voiceover); headings, lists, and code are on-screen. Two directives override the split.

::: scene { id=intro, layout=title, transition=fade }
# autopresenter
Plain paragraphs become the spoken narration and the speaker notes.

::: show {.step}
- Shown, revealed on click
- Not spoken
:::

::: say
Spoken, never shown on screen.
:::
:::

Full grammar: specs/composition-format.md.

The design system

The DesignSystem is Markdown + YAML frontmatter (design-system.md) — typed tokens and short verbal config in the frontmatter, the messaging hierarchy and copy in the body. Tokens are data (path-patchable, schema-validatable); copy is prose. The same parser and patch protocol as the brief and the composition.

---
visual:
  color: { brand: "#7c6cff", accent: "#3ee6c0", bg: "#080a14" }
  type:  { display: "Inter Tight", body: "Inter", scale: [13,15,18,23,32,46,72] }
verbal:
  taglines: ["Brief in. On-brand everything out."]
---
# Messaging
## Product team
**Every asset, on brand, in one pass.**

Reference: specs/design-system.md.

CLI

autopresenter render <deck.md> --target slides [-o <dir>] [--ds <design-system.md>]
autopresenter render <deck.md> --target video  [-o <out.mp4>] [--speaker asteria] [--duration 90]
autopresenter present <deck.md> [--port 4321]      # live hot-reload preview

Every command is headless with JSON output and has MCP parity, so an agent loop can go brief → design system → artifacts with no GUI in the path.

The narrated video

The video target renders each scene to an SVG poster (rsvg-convert, no browser), narrates it with Cloudflare Workers AI — Deepgram Aura-2 (lossless WAV), and muxes with ffmpeg. --duration time-fits the result (pitch-preserving). Post-process guards keep text off the footer, prevent blank/black frames, and normalize audio. Code blocks render as code panels; the stat layout renders as an infographic.

• Voice: --speaker — asteria (default), luna, orion, zeus, arcas, …
• Needs system binaries ffmpeg + rsvg-convert (the slides target stays zero-dependency).

Architecture

Two intermediate representations sit beneath the artifacts. A DesignSystem — visual tokens plus verbal identity, generated from the brief — is the canonical creative IR every artifact inherits, so brand is applied by construction rather than re-specified per tool: change a token or a tagline and every artifact re-derives. Beneath the creative layer, a composition lowers to a SceneGraph, a target-agnostic structural IR, and each output is a projection of it. Slides project straight to a self-contained HTML deck; video and GIF project to an EditList — timed clips, camera moves, and narration cues — rendered by the compositor; the landing page is a sibling web projector that bypasses the SceneGraph entirely, because a site is not a timeline.

The default pipeline is local-first and runs fully offline. An opt-in Cloudflare cloud tier is additive: cross-device and teammate sync via content-addressed objects on R2, durable multi-step processing on Workflows, and a single egress for cloud models through AI Gateway. The local store stays the source of truth and nothing requires the network. The native macOS render core, the ComfyUI-style production graph, and the full data model are detailed in the PRD.