Ai Interactive Learning Agent

AI Interactive Learning Agent

[](https://github.com/MagicianDu/ai-interactive-learning-agent/actions/workflows/ci.yml)

AI Interactive Learning Agent turns technical sources into visual, interactive, feedback-rich learning experiences.

The goal is not to convert books into web pages. The goal is to reconstruct knowledge into lessons that help learners build durable mental models through visual structure, learner action, feedback, misconception checks, and transfer tasks.

The current product form is a React/Vite Web Deck operated through a default learner MCP profile. The long-term product shape is an AI-native learning system that combines:

• MCP tools for stable local capabilities such as source intake, course

creation, publishing, preview, feedback revision, and export.

• Skills for Codex, Claude, OpenClaw-style clients, and future agent runtimes

to operate those tools through natural language.

• Structured lesson and course-pack data that can later power canvas maps,

playgrounds, tutor mode, teacher mode, and assessment mode. These modes are experimental and hidden from the default learner path.

The Web Deck supports multiple authoring intents. The default is build_mental_model; learners may also request professor_lecture_deck when they want a university or graduate lecture-style course deck. This still outputs a Web Deck, not PPTX or slide files.

Generated learner-facing content is Chinese-first by default.

Status

This repository is in OSS alpha.

What works today:

• React/Vite learning workspace with side navigation, web-deck lessons, course

packs, learner feedback, source evidence, and project-library view.

• Structured lesson and course-pack registries.
• Local MCP server with a default learner profile plus explicit authoring and

operator profiles.

• Source-backed Codex-authored flow using public mock fixtures.
• Feedback revision and static-course export paths.
• Stable CI gate and slower source-regression gate.

What is intentionally still evolving:

• Production-grade source parsing for every document type.
• High-quality AI-authored content generation across full books, papers,

patents, blogs, and documentation sets.

• Hosted multi-user service, accounts, sharing, telemetry, and billing.
• Rich playground/tutor/teacher modes; these are experimental until they

improve the core learner loop.

Quickstart

Requirements:

• Node.js 22 or newer
• npm

Install and start the local app:

npm ci
npm run dev

Open the Vite URL printed by the dev server, usually:

http://localhost:5173/

The default app opens a public mock course pack:

智能体工作流公开示例：课程包

Verification

Run the stable open-source gate:

npm run test:ci

This runs:

• TypeScript typecheck
• ESLint
• Stable unit/component tests
• Production build
• MCP/skills bundle check

Run the slower source-regression gate when changing source ingestion, grounding, planning, or MCP learner flows:

npm run test:regression
npm run seed:check

For local MCP installation checks:

npm run bundle:check
npm run codex:mcp:check

Before a public beta checkpoint, run the full local release gate:

npx playwright install chromium # first time on a machine
npm run release:check

release:check runs the stable CI gate, source regression, seed/MCP checks, Codex MCP config check, and a Playwright smoke that verifies the default sample course, a generated #/preview/ route, and next-page navigation.

Public Fixtures

Open-source demos should use public mock sources under:

examples/sources/

Current fixtures:

• examples/sources/agent-workflow-notes.md
• examples/sources/talker-reasoner-architecture.md

Do not commit private books, papers, patents, blogs, notes, generated private lessons, or local machine paths. Runtime artifacts belong under ignored runs/.

MCP Server

List available learner-profile MCP tools:

npm run mcp -- --list-tools

Advanced profiles are explicit:

npm run mcp -- --list-tools --profile authoring
npm run mcp -- --list-tools --profile operator

The default learner profile is intentionally small:

prepare_learning_course -> publish_learning_course -> get_learning_preview -> revise/apply_revision -> export

Prepare a learner-first course request against the public mock source:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"public-mock-smoke","version":"0.0.0"}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"learning_agent.prepare_learning_course","arguments":{"request":"请把这份资料生成中文学习网页，先给总览课，再按核心 topic 拆课，每个单元 8 页，面向有编程基础的中文学习者，教学难度为大学高年级/研究生课程。","runId":"public-mock-smoke","sourcePath":"examples/sources/agent-workflow-notes.md","sourceKind":"book","audience":"有编程基础但缺少系统心智模型的中文学习者","difficultyLevel":"upper_undergraduate_or_graduate","unitPages":8,"strategy":"overview_plus_topic"}}}' \
  | npm run mcp

After Codex or Claude authors and publishes the returned coursePack and lessons, run the app and open:

http://127.0.0.1:5173/#/preview/public-mock-smoke

Learner-facing tools:

• learning_agent.prepare_learning_course
• learning_agent.list_learning_projects
• learning_agent.archive_learning_project
• learning_agent.publish_learning_course
• learning_agent.get_learning_preview
• learning_agent.revise_learning_course
• learning_agent.apply_learning_revision
• learning_agent.export_learning_course

Advanced authoring profile adds source context and quality-comparison tools such as create_learning_project, get_authoring_context, compare_authoring_quality, create_quality_revision, and generate_grounded_course.

Operator profile adds explicit expert review and debug tools, including plan_run, init_from_plan, beta_status, read_artifact, approve_gate, run_course, and promote_units.

Default learner-facing clients should not ask users to approve internal source-map, concept-map, or curriculum-plan artifacts. Use those gates only when the user explicitly asks for expert/operator mode.

Try With Codex

Install the local MCP config and skills bundle for Codex:

npm run codex:bundle:install
npm run bundle:check
npm run codex:mcp:check

The relevant local skills live under skills/:

• learning-agent-operator
• source-to-course
• learner-feedback-revision

Copyable trial prompts are in:

docs/runtime/codex-user-trial-script.md

The intended natural-language flow is:

1. Clarify the learning goal and source constraints in a few questions.
2. Call learning_agent.prepare_learning_course.
3. Let Codex or another capable agent author the Chinese course pack and

lessons.

1. Call learning_agent.publish_learning_course.
2. Call learning_agent.get_learning_preview.
3. Use revise_learning_course and apply_learning_revision for learner

feedback.

1. Use export_learning_course for a shareable static artifact.

Default learner-facing answers should return the preview URL and compact qualityReport summary. They should not ask learners to approve internal source-map, concept-map, or curriculum-plan artifacts.

Project Structure

src/
  components/          Reusable deck, visual, interaction, assessment, course UI
  course-packs/        Course-pack registry and public examples
  lessons/             Lesson registry and public examples
  product/             Learning workspace shell and routing
  renderers/           Web deck, canvas map, and product renderers
  schemas/             Structured lesson and course-pack types
tools/
  agent-runtime/       Local generation runtime and CLI
  mcp-server/          stdio MCP entrypoint
skills/                Agent operating skills
docs/
  product/             Current product definition and core learner loop
  runtime/             MCP, Codex, source grounding, preview, and export guides
  archive/             How to interpret historical planning documents
  superpowers/         Historical specs and implementation plans from development
examples/sources/      Public mock source fixtures

Core Scripts

npm run dev                 # start the local app
npm run build               # typecheck and build the app
npm run typecheck           # TypeScript only
npm run lint                # ESLint
npm run test                # stable unit/component tests
npm run test:ci             # full stable OSS gate
npm run test:regression     # slower source/MCP regression gate
npm run mcp -- --list-tools # inspect MCP tools

Contributing

See CONTRIBUTING.md.

Before opening a pull request:

npm run test:ci

When touching source-grounding or MCP flows, also run:

npm run test:regression

Before a public beta release candidate, run:

npx playwright install chromium
npm run release:check

License

MIT. See LICENSE.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: MagicianDu
• Source: MagicianDu/ai-interactive-learning-agent
• License: MIT

Install and usage instructions live in the source repository linked above.