Memora

Memora MCP

[License: MIT] [MCP Apps] [TypeScript] [React]

Turn any conversation into an interactive study session, right inside Claude Desktop. Ask Claude to make a deck and it generates the cards, then renders them as an inline review you can click through: flip flashcards, answer multiple-choice quizzes, and fill in cloze blanks. Every answer updates a spaced-repetition schedule, and your results flow back to Claude so it can drill what you missed.

Keywords: Model Context Protocol, MCP server, MCP Apps, Claude Desktop, flashcards, quizzes, spaced repetition, SRS, FSRS, Anki alternative.

Built on the MCP Apps extension (SEP-1865): core MCP spec 2025-11-25 plus the Apps extension 2026-01-26.

[Three ways to study, rendered inline in Claude]

Features

Three ways to study, all rendered inline and all graded the same way:

• Flashcards: click to flip, then grade Got it / Missed it.
• Multiple-choice quizzes: pick an option and the correct/wrong answers reveal instantly.
• Cloze (fill-in-the-blank): the blank reveals in place so the sentence stays intact. Write a blank as [...] in a card front and it becomes a cloze card automatically.

Spaced repetition: every grade updates a per-card FSRS schedule (via ts-fsrs), so weak cards resurface sooner and due cards come first. Ask "what's due today?" for a summary across all your decks.

Organize with categories: name decks with :: to nest them (e.g. LLM::Attention). Browse the category tree, or the same tree as an interactive mind map, and study a whole branch in one merged session. Quiz decks are flagged with a badge.

Manage your decks: create, append to, edit, rename, and delete decks and cards without leaving the chat. Pass reverse when creating a deck to also drill each card back-to-front (handy for vocabulary).

Just JSON: decks live in data/decks.json, read live on every call. Hand-edit them or let Claude build them. No database, no external service.

[Browse decks as a category tree or a mind map]

The tools

Only review_deck and study open the inline flip-card UI; everything else returns a text summary, so editing or creating in chat no longer pops the widget.

Card generation follows Memora's quality rules (atomic single-concept cards, 1-5 word answers, active recall, unambiguous, no answer leakage, application over definition), based on Wozniak's 20 Rules of Formulating Knowledge. New cards are linted on write and flagged (never blocked) so Claude can revise them.

How it works (MCP Apps)

A tool declares a ui:// resource. When Claude calls the tool, the host (Claude Desktop) fetches that resource, renders its HTML in a sandboxed iframe, passes the tool result to the UI, and the UI talks back to the host over JSON-RPC (to grade cards and report progress to the model).

[The MCP Apps round trip]

Quick start

Prerequisites

• Node.js 20.11+

Connect to Claude Desktop

Open Settings > Developer > Edit Config and add Memora under mcpServers:

{  "mcpServers": {    "memora": {      "command": "npx",      "args": ["-y", "@servation/memora-mcp", "--stdio"]    }  }}

Then fully quit Claude Desktop (from the system tray) and relaunch. memora appears under Settings > Developer, pre-loaded with a few sample decks.

Your decks are stored in ~/.memora/decks.json (override the path with the MEMORA_DECKS environment variable). Hand-edit that file or let Claude manage it.

git clone https://github.com/Servation/memora-mcp.gitcd memora-mcpnpm installnpm run build

{  "mcpServers": {    "memora": {      "command": "node",      "args": ["C:\\path\\to\\memora-mcp\\dist\\main.js", "--stdio"]    }  }}

Try it (in a Claude Desktop chat)

• review my World Capitals deck
• make me a deck of 10 Spanish travel phrases
• make a reversible deck of 10 Spanish words (drills both directions)
• make a fill-in-the-blank deck about the water cycle (cloze cards)
• quiz me with 10 multiple-choice questions on the solar system
• turn what we just discussed into a deck called "Photosynthesis"
• study my LLM category
• what's due today?

Tech stack

• Server: TypeScript, @modelcontextprotocol/sdk plus @modelcontextprotocol/ext-apps, stdio transport (Streamable HTTP also available).
• UI: React plus Vite, bundled to a single inlined HTML file via vite-plugin-singlefile.
• Scheduling: ts-fsrs (FSRS).
• Runtime is plain node once built (no bun or tsx needed).

Deck format

{  "Deck Name": [    { "front": "Capital of France?", "back": "Paris" }  ]}

• Stored in ~/.memora/decks.json when installed (or data/decks.json from source); override with MEMORA_DECKS.
• Read live (mtime-cached). create_deck, create_quiz, and grade_card write here atomically.
• A quiz card adds "options": ["...", "..."]; its back is the correct option.
• A cloze card writes the blank as [...] in the front, with the hidden term as the back.
• Cards gain a due date and an FSRS srs block as you review them; cards without them are treated as new.
• Keep it valid JSON, or the server falls back to a built-in default deck.

Project structure

memora-mcp/├── server.ts            # MCP tools + the ui:// resource├── decks.ts             # data model, decks.json storage, result builder├── scheduling.ts        # FSRS scheduling + review ordering├── main.ts              # entry: stdio (Claude Desktop) or Streamable HTTP├── mcp-app.html         # UI entry HTML (bundled by Vite)├── src/│   ├── mcp-app.tsx      # review orchestrator (flip / quiz / cloze, grade -> model)│   ├── deck-lib.tsx     # tree, mind map, card list, quiz/cloze views, helpers│   ├── mcp-app.module.css│   └── global.css       # host theme variable fallbacks (light/dark)├── data/decks.json      # editable decks, read live├── media/               # README images├── vite.config.ts       # single-file bundle config└── tsconfig*.json

Development

npm run dev        # vite watch (UI) plus tsx server on http://localhost:3001/mcpnpm run typecheck  # tsc --noEmit

For fast local iteration you can also run the app against the MCP Apps reference host (basic-host) from the ext-apps repo.

Roadmap

See TODO.md for the backlog: npm + MCP Registry publishing, cross-client hosting (Streamable HTTP), tests around parsing and scheduling, and a real screen-capture demo GIF.

License

MIT