Presentations

Most AI presentation tools generate forgettable decks with stock-art-looking images, blue-themed gradients, and bullets that read like keyword soup. They also live in someone else's cloud, behind someone else's pricing page.

Meggy generates presentation-grade decks from a single prompt — and does it locally, on your machine, with your vault as the source of truth.

What you get from one prompt

"Create a 5-slide presentation on the benefits of remote work."

You get a .pptx file under ~/Meggy/Presentations/ that opens in PowerPoint or Keynote and looks like a senior consultant built it:

That's the baseline. The differentiators are below.

Vault-grounded, zero-hallucination decks

Meggy's vault-grounded mode produces citation-backed decks. Ask:

"Build me a 10-slide deck on Q3 revenue trends from ~/Vault/Reports."

Meggy searches the vault, drafts the deck, and refuses to put a number on a slide unless it traces to a vault chunk. A built-in citation validator scans every numeric stat (13%, $50K, 1,200) and blocks export if any are uncited. The final slide is auto-generated as a sources slide listing every citation.

No SaaS competitor can ship this. Vault-grounded decks need local file system access, persistent identity, and an embedding-search pipeline — they need to become Meggy.

Scheduled, recurring decks

Want a weekly board update generated automatically from your vault every Monday at 8am, emailed to your team? One built-in agent handles this:

agent: com.meggy.weekly-deck-builder
schedule: "0 8 * * 1"
topic: "Weekly board update"
sourceFolder: "/Vault/Reports"
themeName: "midnight-executive"
deliveryChannel: "email"
recipient: "team@yourcompany.com"

Each fire pulls the week's data, expands a saved DeckSpec template with fresh figures, exports as PPTX, and sends through the configured channel. This is the kind of feature that requires a local agent runtime — Gamma's API needs a human operator behind every call.

Voice-driven editing

When the voice channel is active and a deck is open, voice transcripts route to a structured edit parser. Say:

"On slide 3 swap the photo for a bar chart, and make slide 5 more skeptical, then add a slide about Q4 outlook."

Meggy parses this into three edit ops, applies them, and re-exports the deck. Multi-turn voice editing of a live deck is impossible for SaaS products — they don't own the desktop microphone.

Multi-agent quality pipeline

For decks that matter, Meggy ships a four-agent pipeline:

  1. Researcher — gathers facts from vault + web.
  2. Writer — drafts the DeckSpec outline with appropriate slide kinds.
  3. Designer — picks layouts, themes, and visuals (icons over AI images, real charts over fake-chart photos).
  4. Critic — runs the 16-item visual QA rubric, applies fixes, re-exports until the score reaches 9/10.

The pipeline is forkable: rebind the critic to a local Ollama model, swap the designer's prompt, or freeze a custom rubric. SaaS tools have a single "Generate" button. Meggy has a teachable pipeline you can edit.

Confidential mode

For legal, medical, or sensitive decks: enable confidential mode on the persona. Every model call resolves to a local provider (Ollama / LM Studio). Web search and cloud image-gen are refused. The renderer fail-fasts if a needed role can't resolve locally — no silent leakage to the cloud, ever.

Brand kit per persona

Each persona carries its own brand kit (logo, primary / secondary colors, heading + body fonts, voice tone). When the persona generates a deck, the brand kit overrides the chosen theme's accent — bg and structural fields stay with the theme. Switch personas → switch brand instantly.

postExport hook

Every successful export fires a postExport hook. Wire your own automation:

Meggy's hook engine handles webhooks, scripts, agent chains, notifications. You don't need to wait for us to build the integration you need.

What's deferred (and why we're transparent about it)

v1.0 deliberately defers stock photo providers (Pexels + Unsplash) because the registration and attribution overhead outweighs the v1.0 ROI. The icon-first visual strategy plus a hardened AI-image prompt builder (with multimodal rubric retry) cover most decks well. We pick up stock providers when telemetry shows the AI-image branch underperforming on photographic slides.

Also deferred: real-time multi-user collaboration, infinite-template libraries, free-tier viral strategies. These are categories we deliberately don't compete in — see the implementation plan for the rationale.

Try it

Open Meggy and ask:

Create a 5-slide presentation on [your topic].

Then check ~/Meggy/Presentations/. The DeckSpec is also persisted under .specs/<deckId>.deck.json — open it, tweak it, ask Meggy to re-export. That's the v1.0 loop, and it's the foundation for everything else.