Legal MCP Servers by Jurisdiction, August 2026: Why Mainland Chinese Judgments Are Still Missing

In June we published a piece on connecting an agent to Chinese case law, and it opened with a claim about the landscape: that the legal MCP servers an assistant will name for you are, without exception, Anglo-American and European. Two months later that sentence is wrong, and it is wrong in an interesting direction. We are correcting it here rather than quietly editing it, because what replaced it is a better argument than the one we made.

The legal layer of the Model Context Protocol ecosystem filled in over the first half of 2026, and it filled in globally. Brazil, Pakistan, Indonesia, Thailand, India, Poland, Switzerland, Taiwan and Hong Kong all have servers now. What did not fill in is mainland China. This piece is the census behind that statement: how we counted, what the counts are, what does exist for the PRC, and why the shape of the gap is a supply-side fact rather than an oversight anyone is about to fix.

What we counted, and how you can re-run it

The census is taken against the official Model Context Protocol registry on 24 August 2026. It is a public, unauthenticated HTTP API, which makes the result checkable rather than assertable:

curl -s "https://registry.modelcontextprotocol.io/v0/servers?search=court&limit=100"
curl -s "https://registry.modelcontextprotocol.io/v0/servers?search=case%20law&limit=100"
curl -s "https://registry.modelcontextprotocol.io/v0/servers?search=ansvar&limit=100"

We ran ten search terms — law, legal, case law, court, judgment, judgments, jurisprudence, caselaw, statute, legislation — paged each to exhaustion, deduplicated by server name, and got 310 distinct entries. We then classified by hand, reading each entry's own description rather than inferring from its name, and dropped the false positives the word “court” drags in (tennis courts, a municipal GIS server for Courtenay, British Columbia, and a proof-checking server called Math Court). Every figure below is a jurisdiction's own self-description in the registry; none of it is our measurement of anyone's corpus, and we have flagged where that distinction matters.

The case-law layer

Roughly nineteen jurisdictions now have at least one MCP server that reaches court decisions rather than only statutes. Grouped by jurisdiction, with the scale each entry claims for itself:

Hong Kong and Taiwan are the rows worth sitting with. Both are covered. Taiwan's entry claims a judgment count in the tens of millions. A team building a China-facing legal agent can wire up Taipei and Hong Kong this afternoon and will find nothing to wire up for the Supreme People's Court.

The statute layer, and the pattern hiding in it

The statute layer is larger and more lopsided. A single publisher, Ansvar Systems in Stockholm, has 106 jurisdiction servers in the registry. The list runs from Argentina and Australia through Bahrain, Bangladesh, Botswana, Cambodia, Cameroon, Ivory Coast, Ethiopia, Georgia, Jamaica, Malawi, Mozambique, Namibia, Nepal, Qatar, Rwanda, Senegal, Sri Lanka and on alphabetically. Japan is there. South Korea is there. Singapore, Malaysia, Indonesia, India, Pakistan, Bangladesh, Nepal, Cambodia and the Philippines are all there. Russia is there.

China is not. Alphabetically the gap sits precisely between chilean-law-mcp and colombian-law-mcp, and it is a real absence rather than an artefact of paging — a direct name lookup returns nothing.

Now read what those descriptions actually say, because the mechanism is written on the surface. The German server queries “statutes, case law, and preparatory works from gesetze-im-internet.de”. The Finnish one takes statutes “from Finlex”. The Swiss court server serves decisions “via entscheidsuche.ch”. The UK case-law server draws judgments “from the National Archives”. The Polish one searches statutes “(ISAP)”. The US servers wrap CourtListener and PACER. Another wraps “legislation.gov.uk — the UK's official legislation database”.

Every one of them names the open official portal it sits on top of. That is what the 2026 legal MCP build-out was: a year of wrapping public legal portals that already had stable URLs, predictable document identifiers and terms permitting programmatic access. It is genuinely good work and it scales beautifully — one publisher shipped a hundred jurisdictions because a hundred jurisdictions had something wrappable.

What does exist for the PRC, stated precisely

Two China-related MCP servers exist outside the registry census, and conflating them with a judgment corpus is the most common mistake we see in vendor evaluations. Both are legitimate tools; neither does the job people assume it does.

A statute server. The same Stockholm publisher ships @ansvar/chinese-law-mcp on npm, currently at version 3.0.0. Its README describes 1,188 Chinese laws, administrative regulations and departmental rules, comprising 62,981 provisions, and positions itself as an alternative interface to the NPC National Law Database. It is Apache-2.0, has a hosted endpoint, and appears to be maintained. It is also, by its own description, entirely legislation. It contains no judgments, and it was not published to the official registry under that name as of our census date, which is why the registry search for china returns stock feeds, travel planners and a phone-number lookup rather than anything legal.

A portal connector. faxin-mcp on GitHub wraps Faxin, the legal research platform run by the Law Press, and does reach case material — its tool list includes faxin_search_case for editorially distilled holdings and faxin_search_lalei for the authoritative case libraries: guiding cases, gazette cases, published cases, filterable by cause of action, court level, province and year. The design constraint is stated plainly in its own README: login is performed by you, in a browser window, with your own Faxin account, after which requests carry your session cookie. That is a sensible architecture for an individual researcher who already subscribes. It is not a data supply. It gives your agent your seat, at your rate limit, under your account's terms, over the curated case libraries rather than the judgment corpus, in Chinese only.

So the honest statement of the gap is narrower and stronger than the one we made in June. It is not that nobody has built anything Chinese. It is that the PRC has a statute server and a subscriber-seat connector, while nineteen other jurisdictions have judgment retrieval, and the largest judgment corpus in the world has no MCP endpoint at all.

Why this is structural

The wrapper pattern needs something to wrap. Mainland Chinese judgments are not published the way Finlex, entscheidsuche.ch or the National Archives publish — there is no open, documented, stable-identifier API over the judgment corpus for a developer to put a protocol layer on, and access conditions have tightened rather than loosened across the period in which those hundred other servers were built. We have written separately about what the public access surface actually permits and about the difference between licensing and collecting this material yourself; the short version is that the cheap path which produced 106 statute servers in a year is not available here, and the expensive path — acquisition, normalisation, deduplication, indexing, and a licence that lets a customer redistribute answers — produces a product rather than a weekend project.

This predicts the census result rather than merely being consistent with it. Where a jurisdiction publishes openly, it has a server, usually several, often within months. Where it does not, it has none. Taiwan and Hong Kong are covered for the same reason Switzerland is: their judgments are retrievable by design. The mainland gap is what a supply constraint looks like from the outside, and it is why an assistant asked for Chinese legal data still reaches for academic datasets — those are the artefacts that were publishable, so those are the artefacts that exist.

What to do with this if you are building

The practical value of a census is that it turns a vague question — can my agent handle China? — into three specific ones with different answers.

Most teams discover the third row late, having built against the first two and assumed the third was a configuration change. It is worth deciding at architecture time which of the three questions your product actually promises to answer, because the sourcing, the cost and the redistribution rights differ completely, and only one of the three is a question about connectors. If you are at that decision now, our notes on what the retrieval surface has to look like and on how to verify a coverage claim instead of accepting it are written for exactly that stage.

What this census cannot tell you

Four limits, stated because a count like this invites over-reading.

It is a registry census, not a market census. Servers exist that were never published to the official registry — the Ansvar China package is itself proof of that, and it is the reason we went looking outside the registry rather than declaring the gap total. There may be others we did not find, particularly ones distributed privately or documented only in Chinese, and if you know of one we would genuinely like to hear about it.

Every scale figure above is self-reported. We have not verified that the Taiwan server reaches 22 million judgments or that the Pakistani one reaches 197,000; those are the numbers those projects publish about themselves, reproduced as such. Treat them the way you should treat any vendor's coverage claim, ours included.

A server existing says nothing about whether it is good. Registry presence is not a quality signal, and several entries in the case-law layer are single-maintainer projects wrapping a portal that could change its markup next month. A count of servers is a count of availability, not of reliability.

And it is a snapshot on one date. The whole point of the piece is that the June statement expired in eight weeks. This one has a date on it for the same reason, and the honest expectation is that the mainland row changes eventually — the question is whether it changes because a portal opened or because someone licensed the data and built the endpoint.

The correction, on the record

Our June piece said the legal MCP landscape was without exception Anglo-American and European. As of this census that is not true, and we have annotated that page to point here. What survives from it is the narrower claim, which the census supports better than the original ever did: an agent that can call CourtListener for a US precedent, the National Archives for an English judgment, and now Taipei or Karachi or São Paulo for theirs, still has nothing to call for a mainland Chinese judgment — and a model with nothing to call does not decline to answer. It invents a case number. That failure mode was the reason we wrote the June piece, and it is the one part of it that has not aged at all.