learndeck
Enables AI-assisted learning through structured courses with Socratic guidance, evaluating user answers against rubrics and managing learning progress locally.
README
LearnDeck — your app for learning with AI
LearnDeck is a forkable, local app for learning with AI without outsourcing your thinking. It pairs structured, project-based courses with the learner's own workspace. An MCP-connected AI is a calm, Socratic guide: it asks the next useful question, evaluates a visible answer against an author-written rubric, and never writes the learner's solution for them. Answers, evidence, feedback, and learning records stay in local SQLite.
Courses live in the public catalogue at learn-deck/courses and sync into the app; the v0.1 flagship is DDD and Hexagonal Architecture — six to eight hours of Node.js + TypeScript in satisfying 45–60-minute building sessions. This repository bundles only a small format example pack for documentation and development. The format remains course-agnostic while the default catalogue grows deliberately and stays curated for quality. See the product position and catalogue quality rubric.
Run locally
Requires Bun 1.3 or newer. Check your version first:
bun --version
From this clone, install dependencies and start the app:
bun install
bun run app
If startup reports EADDRINUSE or that port 3030 is already in use, stop the
other LearnDeck process, or run PORT=3031 bun run app and open the URL
LearnDeck prints.
Open http://127.0.0.1:3030. LearnDeck creates local
progress at .learndeck/progress.db. Set LEARNDECK_DB_PATH to use a separate
database.
What stays local
- Progress is SQLite at
.learndeck/progress.db;LEARNDECK_DB_PATHoverrides it. - Connecting Codex targets
~/.codex/config.toml, Cursor targets~/.cursor/mcp.json, and Claude Code targets~/.claude.json. - A connection adds exactly one
learndeckMCP entry to the selected guide. Disconnect removes only that entry throughDELETE /api/integrations/:id/connect. - Per-path progress can be exported with
GET /api/paths/:id/exportor reset withDELETE /api/paths/:id.
The primary catalogue is the public GitHub course repository; while no
repository is configured, only the bundled example-course format pack
loads, which keeps development working offline. To select the public catalogue, copy the release
configuration before bun run app:
cp .env.example .env
The release configuration selects github:learn-deck/courses@main. When the
learner clicks Start Now, LearnDeck syncs only Markdown under courses/
and references/ into its local cache. If GitHub is unavailable, it uses the
last complete cache. A fork can set a different
LEARNDECK_COURSE_REPOSITORY=github:your-org/courses@main. To publish a
course of your own, follow public course
distribution; for failures, see
troubleshooting.
On first launch, LearnDeck:
- detects Codex, Cursor, and Claude Code locally without launching them or reading credentials;
- lets the learner choose any detected Codex, Cursor, and Claude Code hosts, adding only LearnDeck's own MCP entry to those selected and asking for a restart;
- lets the learner choose an active guide and switch among connected guides without losing local progress;
- introduces the app, lets the learner explore courses by category, and makes each course's time, level, and outcome visible before entry;
- asks only for the learner's project workspace after they choose a course; and
- displays the next action, sources, questions, answer history, and source-linked agent feedback.
Connection is opt-in. The app never installs an agent, starts a learner's server, runs submitted code, or changes unrelated MCP servers.
Add a course
bun run seed -- api-design-basics "API Design Basics"
This creates a Markdown-only course pack (seeding fails if the course ID already exists, so pick a new one):
courses/api-design-basics/
course.md
modules/00-orient.md
course.md contains course identity, category, tags, duration, outcomes,
prerequisites, and the runtime the app resolves for the learner. Each ordered
Markdown module contains its goal, bounded action, source links, questions, and
author-written evaluation rubrics in YAML front matter plus learner-facing
Markdown. The loader validates local source links are real .md files. See
the course-pack standard for the diagram, contract,
and authoring checklist.
The browser UI is a dark-first learning environment with local theme preference, Focus Mode, section-based progress, and accessible source-rendered lesson blocks. Its maintainable design rules live in the UI system.
Agent integration
LearnDeck exposes one local stdio MCP server. Use the app's Connect button for Codex, Cursor, or Claude Code. Another compatible host can use the equivalent of:
{
"mcpServers": {
"learndeck": {
"command": "/absolute/path/to/bun",
"args": ["/absolute/path/to/learndeck/src/mcp.ts"]
}
}
}
The documented tools list courses, paths, progress, the next activity, reported evidence, and submitted-answer evaluation. See MCP integration.
Verify
bun run verify
Optional: a local macOS app
On macOS you can build a double-clickable LearnDeck.app for your own machine.
It needs swiftc (Xcode Command Line Tools) and Bun:
bash scripts/package-macos.sh
open dist/LearnDeck.app
The script compiles the server into a standalone binary, stages public/,
courses/, and references/ inside the bundle, and compiles a native
AppKit/WKWebView shell (native/macos/LearnDeckApp.swift). Launching the app
starts the server on a free local port and opens a native window; quitting the
app stops the server. The app's data lives outside the bundle at
~/Library/Application Support/LearnDeck/ (progress.db, course-cache/,
and server.log), so rebuilds never touch progress.
This is developer tooling, not a distribution channel: the app is unsigned and not notarized, and the supported install remains cloning the repository. One known limitation: connecting an AI guide from the packaged app writes an MCP entry that points at this repository checkout, so keep the clone in place or reconnect after moving it.
Privacy and scope
The browser binds only to 127.0.0.1. Progress, answers, workspace paths, and
reported evidence remain in the local database and are ignored by Git. By
default, that database and the public-course cache live under .learndeck/;
the packaged macOS app keeps them under
~/Library/Application Support/LearnDeck/ instead. See the local progress
database and SECURITY.md.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。