Xuexi Keben — Chinese Textbook Tutor MCP

A free, read-only, model-provider-independent MCP server for the revised Elementary Chinese Readers / 基础汉语课本 corpus.

• 44 lessons across two volumes
• 320 retrieval-ready textbook chunks
• Focused lesson study, Chinese/pinyin/English search, and grounded quizzes
• Standard MCP Streamable HTTP transport
• No model API, model account, database, embeddings, or server-side student data

Connect in 1-2-3

Share this MCP URL:

https://keben.555420.xyz/mcp

1. Open MCP, Apps, or Connectors in your AI application.
2. Add a remote Streamable HTTP server and paste the URL above. Authentication is None.
3. Enable it in a chat and say:

Teach me Lesson 1 from the Elementary Chinese Readers textbook. Then quiz me one question at a time and explain my mistakes.

The same URL works with any client that supports remote MCP. The client chooses the LLM; this server does not know or care whether it is Claude, ChatGPT, Codex, Gemini, or a local model.

Let an agent guide installation

Send the agent the public home-page URL, without /mcp, and say:

Open this URL, read its agent setup instructions, and help me connect its MCP server. Ask before changing my configuration.

The deployed home page links to /llms.txt, health metadata, and a machine-readable server card. MCP clients intentionally require user or administrator approval before trusting a new server, so a normal web chat cannot silently install one.

Claude

In Claude web or Desktop:

1. Go to Settings → Connectors → Add custom connector.
2. Name it Elementary Chinese Readers and paste the /mcp URL.
3. In a chat, select + → Connectors, enable it, and ask to study a lesson.

In Claude Code, an authorized agent can run:

claude mcp add --transport http --scope user chinese-textbook https://keben.555420.xyz/mcp

ChatGPT

Custom remote MCP apps currently depend on the ChatGPT plan and workspace settings. Where developer mode is available:

1. Open Settings → Apps → Advanced Settings and enable developer mode.
2. Choose Apps → Create, provide the /mcp endpoint, select no authentication, scan the tools, and create the app.
3. Enable the app in a new chat.

On managed university workspaces, an administrator may need to create or approve the app once for the class.

Deploy once for the whole class

The repository is ready for a free Cloudflare Worker deployment.

Or deploy from a terminal with Node.js 22 or newer:

npm install
npm run check
npm run deploy

Cloudflare will print a URL ending in workers.dev. Append /mcp and test it before sharing.

For local development:

npm install
npm run dev

The MCP endpoint is http://localhost:8787/mcp. Test it with:

npx @modelcontextprotocol/inspector@latest

MCP capabilities

The server also exposes lesson-index resources and reusable study_lesson and quiz_lesson prompts. Quiz questions are created by the student's chosen LLM from returned textbook material, keeping this server model-neutral and free to operate.

Project layout

• src/ — stateless MCP Worker, retrieval, tutoring rules, and public setup page
• textbook/chunks.jsonl — canonical retrieval corpus
• textbook/lessons/ — human-readable lesson files
• scripts/build_mcp_corpus.mjs — creates the compact Worker data bundle
• scripts/validate_corpus.py — validates the canonical corpus
• scripts/validate_mcp_server.py — validates the generated MCP data

The OCR-derived corpus carries confidence values and quality flags. The MCP server returns those flags with source page metadata so tutors can communicate uncertainty instead of guessing.