A Free Chinese Dictionary API and MCP Server for AI Agents

A chat assistant can already write you a set of Chinese flashcards. What it does not have is a fixed word list to check against, so “this word is HSK 3” is only as good as its memory. I wanted a source it could ask instead, so I put the HSK Studio dictionary behind an API and an MCP server. It went live this week and needs no key. I do not yet know whether anyone will call it.

I’m Giovanni, and I build HSK Studio under Fulin Labs. The data is the dictionary and the graded readings the site already uses. Nothing was added for the API.

What it answers

Five questions, each one a GET request:

• What does this word mean, and which HSK level is it?
• Which words belong to HSK level N?
• Which graded readings exist?
• What is the text of one reading?
• What HSK level does this passage need?

The word lookup takes Chinese, pinyin or English. Here is the real response for 咖啡:

GET https://hsk.fulinlabs.com/api/agent/word?q=咖啡

{
  "query": "咖啡",
  "match": {
    "word": "咖啡",
    "pinyin": "kāfēi",
    "meaning": "coffee",
    "partOfSpeech": "noun",
    "hskLevel": 2
  },
  "related": [ ... ],
  "url": "https://hsk.fulinlabs.com/dictionary?utm_source=agent&utm_medium=api&utm_campaign=agent-api"
}

The level check takes a passage and reports the HSK level whose words cover it. For 我要一杯咖啡。 it returns HSK 2 with 100 percent coverage, 5 unique words and 咖啡 as the hardest one. Level 1 alone holds 369 words, and words?level=1&limit=20 pages through them.

Using it from an MCP client

The MCP endpoint is https://hsk.fulinlabs.com/api/agent/mcp. It takes POST requests with JSON-RPC and keeps no session. The tools are lookup_word, list_words_by_level, list_readings, get_reading, check_text_level and get_level_check_link. Point any client that supports remote MCP servers at that URL.

A plain HTTP client works too. The index at /api/agent lists the endpoints, and /llms.txt describes them for agents that read that file.

Where it stops

It returns no audio. Speech runs in the browser, so every result links to the HSK Studio page for that word, where the learner can hear it and read example sentences.

Poems and classical texts come back with an HSK level of null. No level covers 90 percent of their words, and a guessed level would be worse than none. The reading list shows null for the poems, for example 静夜思 and 春晓.

Page sizes are capped at 100 words, and a level check reads at most 5,000 characters.

Why every result has a link

Each response includes a url on hsk.fulinlabs.com with utm_source=agent. The assistant is meant to show it to the learner. That gives me one number I can read: how many visits come from an agent that used the API. Until I see that number move, “agents will use this” stays a guess.

The reverse is also true. If you build a Chinese study tool and want a word list with HSK levels that you did not have to compile, the API is there. Tell me what is missing from it.