proofread.law

proofread-mcp

An MCP server for proofread.law. It lets Claude Desktop, Claude Code, Cursor and the OpenAI Agents SDK check case citations before a draft is filed: about 10 million US court opinions, and 1.1 million Swiss decisions (BGE/ATF/DTF, Federal Supreme Court dockets, the federal courts and all 26 cantons). The jurisdiction is detected from the draft, and a Swiss report comes back in the draft's language: German, French or Italian.

proofread.law checks each citation against an open register of about 10 million court opinions (CourtListener bulk data). Every result says what was checked, what was found, and what the register cannot see.

What it does

Fourteen tools:

The compact result of a check is capped at about 12,000 characters; when a long brief has more flagged rows than fit, the text says how many were left out and render_report has them all.

Saved briefs (opt-in)

Nothing is saved unless save_brief or update_brief is called; check_citations and check_document never save anything. A saved brief is stored encrypted in the user's proofread.law account until delete_brief removes it, so the brief tools need an API key (any PROOFREAD_API_KEY, or one from sign_up); without one they answer with a message and send nothing. A server with storage switched off answers storage_off, and the tools say so.

The loop for fixing flagged citations: save_brief once, fix a flagged row in the text (the citation, the case name, the quotation, or take the citation out), call update_brief with the id and the whole edited text, and read what changed: which flags were resolved (flagged before, not flagged now), which are new, and the rows that still need attention. The last 20 versions are kept; get_brief lists them and reads an earlier one. Ids are integers in the API and strings in the tools.

check_citations and check_document read prose: they compare the case name and any quotation with the register. resolve_citation and resolve_citations look the citation string up in the register (the /v1/resolve API) and tell you which case sits there; they do not compare it with the name you have.

Tiers, in the words the tools use:

What it cannot do: resolve Westlaw (WL) or Lexis identifiers, check statutes, regulations or secondary sources, or say whether a case is still good law. Those limits are stated in every tool description and in the coverage statement that comes with every result.

Case suggestions (Switzerland)

suggest_cases takes a Swiss federal statute article (Art. 41 OR, art. 41 CO, Art. 8 ZGB, art. 9 Cst.; lowercase and dotless forms such as art 41 or are read too) or a paragraph that cites articles, and lists the leading Federal Supreme Court cases (BGE) the court cites with it. Federal acts only; cantonal law is not covered. There is no free-text search: a query without an article answers with a message that no article was recognised. The answer is one ranked list in the measured order, each case labelled with its field of law; the domain filter keeps the same order within one field, so a filtered row keeps its overall rank (Rang 7, Rang 9, ...) and a filter note says which field is shown. Each row has the citation, date, field, the decision's language, its rank, the quoted passage (regeste or consideration), any later change of practice (a changed precedent is listed with its flag, never dropped), the decision at the court's site and a link that opens the check on proofread.law. Every string comes back in the query's language (German, French or Italian; lang: "en" for English), and the tool prints it as it is. A suggestion is a case to read: it has not been checked against your sentence, and a row without a flag is not evidence that its practice still holds. To check a citation, use check_citations.

Arguments: query (required, up to 20,000 characters; over 1,500 it is sent as POST), domain (all, civil, criminal, public or social; default all), lang (de, fr, it or en; default the query's language), k (number of rows, 1 to 50; default 10). With a large k, rows past about 20,000 characters of text keep their header line and any practice flag and leave out the passage and links, which stay in the structured result (with the counts per field). During the trial phase it needs a paid-plan or trial API key; otherwise the API answers plan_required and the tool says where to upgrade.

Example call, suggest_cases with {"query": "Art. 41 OR, Art. 97 OR", "k": 3} (dev instance, 2026-09-26; rows 2 and 3 cut):

Gelesen als: Art. 41 OR, Art. 97 OR

1. BGE 146 IV 76, 13.11.2019, Strafrecht, FR, Rang 1 Das Bundesgericht zitiert diesen Entscheid zusammen mit Art. 41 OR; der Entscheid selbst nennt den Artikel nicht. Aus der Regeste: «a) Art. 110 Abs. 1 StGB; Art. 118, 121 Abs. 1 und 382 Abs. 1 StPO; ...» Entscheid öffnen: https://search.bger.ch/ext/eurospider/live/de/php/clir/http/index.php?highlight_docid=atf%3A%2F%2F146-IV-76%3Ade&lang=de&type=show_document Zitat prüfen: https://proofread.law/?cite=Art.%2041%20OR%3B%20BGE%20146%20IV%2076#check

... Die Vorschläge sind publizierte Leitentscheide (BGE). Sie sind danach gereiht, wie oft das Bundesgericht sie zusammen mit dem Artikel zitiert, [...] Ein Vorschlag ist ein Entscheid zum Lesen: Ob er Ihre Aussage stützt, prüft diese Liste nicht, und ein Entscheid ohne Hinweis ist kein Beleg dafür, dass seine Praxis weiter gilt.

With "domain": "civil" the list starts with the filter note and keeps the overall ranks:

Gelesen als: Art. 41 OR

Filter: Zivilrecht. Die Reihenfolge ist dieselbe wie in der ganzen Liste.

1. BGE 132 III 122, 13.09.2005, Zivilrecht, FR, Rang 7 Regeste: «Rechtmässigkeit von im Arbeitskampf eingesetzten Mitteln (Art. 28 BV; Art. 41 und 357a OR). ...» ...

A filter that leaves no rows answers with a message saying so. A later change of practice appears right under the row it concerns, for example Praxisänderung durch BGE 145 III 1, möglicherweise nur teilweise: followed by the regeste sentence that marks it.

Case suggestions (United States, beta)

suggest_cases also takes a sentence from a US draft that states a rule of law (one sentence, or a short paragraph up to 1,500 characters) and lists up to 3 cases whose own text states it. A model reads the opinion of the court in each of the top 10 candidates, and a case is shown only when the model finds a passage in it that states the sentence. Each row has the reference, the court, the year, how the case stands to the court you name (binding here, same circuit, persuasive), any later history the citator found (reversed in part, superseded by statute), the passage the model matched with the paragraphs before and after it (each cut to about 600 characters), a link to the opinion on CourtListener and a link that opens the check on proofread.law. Cases that a later court overruled or reversed are left out, and a note names them.

This is a beta. Every answer starts with a line that says so, with the measured numbers: on the site's own index, of 94 cases shown for 50 sentences from filed briefs, 73 (78%) support the sentence and 4 do not, as read blind by two model readers, not practising lawyers (details). Read the passage before you cite the case. A suggestion is a case to read; to check a citation, use check_citations.

Arguments: court (optional, up to 200 characters) is the court the brief is filed in: a federal court of appeals (9th Cir., Ninth Circuit, ca9), a district court (N.D. Cal., S.D.N.Y., cand), or a state or territory name or postal code (California, CA) for the federal court there. Cases that bind that court come first; without a court, Supreme Court cases come first. State courts are not read yet; a court the API does not read comes back as an error that says what it reads. jurisdiction is auto by default: a Swiss statute article, or German, French or Italian text, is Swiss, and anything else is a US sentence; us or ch sets it. domain, lang and k apply to Swiss answers only, and a US answer is in English.

Input that states no rule of law