Skip to content

AI assistance

DecksMD has optional AI features that are off until you add an API key in Settings → AI. They never run on their own — nothing is sent anywhere unless you explicitly trigger an action.

Generating flashcards with AI in DecksMD
  • Generate — describe a topic (and optionally attach notes or images) and stream new flashcards into an inbox. Keep the ones you want and save them to a new file or append to an existing deck — as header + paragraph, table, or canvas.
  • Refactor — rewrite a single card, or batch-refactor a selection, to improve clarity and fix grammar; or split one dense card into several atomic ones. You review every change before it’s applied.

There are two ways to use the AI features, and both are fully supported — one is not a trial of the other.

Bring your own key — free, and nothing passes through us

Section titled “Bring your own key — free, and nothing passes through us”

Add a key for:

  • OpenAI
  • Anthropic (Claude)
  • Google (Gemini)
  • Any OpenAI-compatible endpoint, including local servers like Ollama or LM Studio.

You pick the provider and model in settings, and the plugin calls that provider directly from your machineapi.openai.com, api.anthropic.com, and so on. No DecksMD server is involved at any point. We never see your notes, your prompts, or even that you ran a generation, because nothing is sent to us to see.

This costs nothing beyond what your provider charges you, and it stays that way. With a local model such as Ollama, your content never leaves your machine at all.

With DecksMD Pro you skip provider setup entirely: you choose a tier — Flash or Thinking — and the model behind it is kept current for you as better ones appear. See Flash and Thinking. Requests go through our backend, which is what makes PDF OCR possible; what that means for your content is set out below.

Free trialDecksMD ProYour own key
Price€0, no card€9.99/monthFree forever — you pay your provider
Volume~100 cards or 10 PDF pages, once~16,000 cards/month, or 3,000 PDF pagesUnlimited
SetupNoneNoneYour own API key
Scanned PDFs (OCR)YesYesNo — text extraction only
PromptingCurated for spaced repetitionCurated for spaced repetitionA short version of ours
ModelsOursFlash & Thinking, kept currentAny supported, including local
Your contentSent with a no-train flagSameNever reaches our servers

The trial’s figure is one budget, not two allowances — 100 cards or 10 pages, or some mix of the two. Full pricing.

Keys are stored locally in ai-keys.json inside the plugin folder and are never written to data.json — so they never leave your device through Obsidian Sync or any vault file-sync.

Nothing is sent to a provider unless you explicitly trigger an AI action. When you do, the request contains only:

  • a built-in instruction prompt describing how DecksMD cards work,
  • your typed instructions, and
  • the relevant content for that action — the card’s fields (for refactor) or your topic/prompt (for generate), plus any notes or images you chose to attach.

The plugin makes no background or telemetry calls. The only network requests are the ones you start — sent directly to the provider you configured, or to our backend if you use DecksMD Pro.

Your notes get read to make cards out of them. They do not get kept.

With your own key, nothing reaches us at all. The request goes from your machine to your provider, under your account. What happens to it after that is governed by your agreement with them.

With DecksMD Pro:

  • We do not store it. Notes, prompts, attached images, transcribed PDF pages and generated cards are held in memory for the length of the request and then gone. There is no table in our database that content can be written to.
  • We send every request with a “do not train” flag set, and route only to providers whose stated policy is not to train on or retain what they are sent. We enforce this on our side, on every request, at the single point where our backend talks out — no code path can skip it. How a provider then handles the request is governed by their own published policies rather than by us.
  • What we do keep is what billing needs: how many tokens a request used, which model served it, and when. Those records identify your account by an opaque id, never by the content of what you generated.
  • Nothing is written to logs. Our operational logs contain token counts, model names and account ids — no note text, no prompts, no email addresses.

The flag is set on every request we make, including the ones to the OCR model that reads your PDFs. If you would rather no third party received your content at all, use your own key with a local model such as Ollama — that path is free and always will be, and nothing leaves your machine.