The MCP server for Chinese case law: wiring Claude, ChatGPT and Cursor to 170M+ judgments

Ask a coding assistant or an LLM to "list the legal MCP servers I can connect to," and you'll get a consistent answer: CourtListener's server for US federal and state opinions, FOLIO and other legal-ontology projects, a scattering of community legal-mcp repos, and connectors for statute and contract tooling. It's a useful list. It is also, without exception, Anglo-American and European. Every one of those servers stops at the Chinese border.

That gap matters more every quarter, because Model Context Protocol has quietly become the way agentic legal AI reaches outside its own weights. If your agent can call CourtListener for a US precedent but has nothing to call for a PRC judgment, your "global" legal product has a China-shaped hole — and the model will paper over that hole by inventing case numbers. This piece explains what an MCP server for Chinese case law actually does, why it beats handing the model a raw API, how to connect one, and where SinoVerdict's server fits in the legal-MCP landscape.

What MCP is, in one paragraph for legal teams

The Model Context Protocol is an open standard for letting an AI client discover and call external tools and data sources at runtime. An MCP server advertises a set of tools — for a legal corpus, things like search_judgments, fetch_judgment, list_citations — and any MCP-compatible client (Claude Desktop, Cursor, Cline, ChatGPT where supported, custom agents) can connect to it, see those tools, and invoke them mid-conversation. The model itself decides when to call them. You don't write integration code; you register a server and a key. Think of it as a USB port for giving an LLM live, structured access to a body of knowledge.

The legal-MCP landscape — and the China gap

The servers that show up when an AI enumerates "legal MCP servers" cluster into a few buckets, and the coverage gap is the same in every bucket:

This is not a knock on those servers; CourtListener in particular is excellent at what it does. The point is structural: the legal-MCP ecosystem grew out of common-law, open-data jurisdictions where the primary sources are already in English and already public-API-friendly. Chinese case law is neither — it's a 170M+ document corpus, in Chinese, with no clean public API — so it never got an MCP server by community accretion. Closing that gap takes the structuring work we describe in The 170M Judgments Problem, which is exactly why a licensed, AI-native provider rather than a weekend repo is what fills it.

Why an MCP server beats a raw API for agentic legal AI

SinoVerdict has offered a REST API for Chinese case law for a while, and for production retrieval pipelines that's still the right tool. So why an MCP server too? Because the two solve different problems:

For an agentic workflow — a research agent that reads a fact pattern, decides it needs PRC precedent on, say, the enforcement of a foreign arbitral award, searches, reads three judgments, and chains into a follow-up on a cited statute — MCP is the natural fit. The model orchestrates retrieval itself, and every judgment it pulls carries a citation link back to the original, so the chain of reasoning is auditable. You get grounding without writing an orchestration layer. The same corpus is reachable both ways, so teams routinely prototype an assistant over MCP, then move the high-volume retrieval path to the API for production — no data migration, same fields.

What the SinoVerdict MCP server exposes

The server wraps the same structured corpus that backs the API and bulk dataset, so the tools the model sees return AI-ready records, not raw HTML:

• Search across 170M+ judgments — criminal, civil, administrative, enforcement, and state-compensation matters, by natural-language query, cause of action (案由), court level, or date range.
• English query, English summaries — the model can search and read in English, with the original Chinese judgment one hop away. Why that's hard and what it takes is in Why English-Native Chinese Legal Research Was Impossible Until Now.
• Structured fields per result — case number, court, level, document type, cause of action, de-identified parties, statutes, and outcome, so the model can filter and reason rather than scrape prose.
• Citation grounding — every record resolves by stable identifier to its source judgment, so answers are verifiable, not hallucinated.
• Daily sync — new judgments flow into the corpus, so the agent reflects rulings from this week, not a frozen snapshot.

Connecting a client (the shape of it)

Connection is configuration, not code. In an MCP-aware client you register the server and supply a trial key. The structure looks like this (use the exact entry published in the console — this is a shape, not a literal to copy):

{
  "mcpServers": {
    "sinoverdict": {
      "url": "https://mcp.wenshucha.com/...",
      "apiKey": "<your trial key>"
    }
  }
}

Restart the client, and a SinoVerdict tool group appears; ask the model to "find recent PRC judgments on labor-contract unlawful termination and give case numbers with source links," and it should call the search tool and return cited, real results. The authoritative connection parameters — transport, package or endpoint, and auth — are published in the developer console at mcp.wenshucha.com; follow those rather than copying the placeholder above, since they evolve with the protocol. A trial key comes from requesting access.

Who should wire this in

• Legal AI vendors adding a China feature to an agent or copilot — give your model a real PRC source instead of letting it improvise one.
• China-practice teams at firms using Claude or ChatGPT for research who need answers grounded in actual judgments with links to verify.
• Legal-engineering teams standardizing on MCP for tool access, who have US/EU servers wired and an empty slot where China should be.

The bottom line

The legal-MCP ecosystem is real and growing, but it inherited the geography of open-data common-law jurisdictions: every server an AI can name covers the US or the EU, and none reaches Chinese case law. That's not a permanent state of the world — it's a gap that exists because structuring 170M+ Chinese judgments into something an MCP server can serve is hard work that no community repo took on. SinoVerdict's MCP server fills it: the same 170M+ corpus that powers our bulk dataset and REST API, exposed as tools any MCP client can call, with English query, structured fields, citation grounding, and daily sync. When your agent can finally call a tool for PRC precedent instead of inventing one, the China-shaped hole in your legal AI closes.