Language Support

Supported File Types

Language Check extracts prose from these file formats using tree-sitter parsers:

Format	Language ID	Extensions	Parser / Strategy
Markdown	markdown	.md, .markdown	tree-sitter-markdown
MDX	(alias)	.mdx	Treated as Markdown
HTML	html	.html, .htm	tree-sitter-html
XHTML	(alias)	.xhtml	Treated as HTML
LaTeX	latex	.tex, .latex, .ltx	tree-sitter-latex
R Sweave	sweave	.Rnw, .rnw	R chunk preprocessing + tree-sitter-latex
reStructuredText	rst	.rst, .rest	tree-sitter-rst
Org mode	org	.org	tree-sitter-org (vendored)
BibTeX	bibtex	.bib	tree-sitter-bibtex
Typst	typst	.typ	tree-sitter-typst (vendored)
Forester	forester	.tree	tree-sitter-forester (vendored)

Prose extraction details

Each language has a custom prose extractor that understands which parts of a document contain human-readable text.

An extractor answers two questions. The first is which AST nodes carry prose, which is a walk over the tree. The second is what sits in the gap between two of those nodes — the markup that decides whether the words on either side belong to one prose block, and which bytes the checker must never see. A language describes its gap syntax once, as a single function that recognizes one token at a byte offset, and both answers are derived from it; see Adding Language Support via the Plugin Path for how to write one. Markdown and HTML take a different path entirely, selecting prose nodes with a tree-sitter query and no gap analysis at all.

• Markdown / HTML — Uses tree-sitter query patterns to select prose nodes: paragraphs, headings and table cells in Markdown, text nodes in HTML. Fenced code blocks, front matter and <script>/<style> bodies are skipped. A Markdown block is handed to the checker whole, so inline code spans and link URLs are skipped by the engine’s own Markdown parser rather than at extraction.

• LaTeX — Tree-walks the AST collecting word nodes from \begin{document} onward. Skips preamble, math environments, verbatim/minted/algorithm blocks, and structural commands (\ref, \label, \includegraphics, etc.). The delimiter-delimited argument of \verb|...|, \verb*|...|, \lstinline!...! and \mintinline{lang}|...| is excluded too, although the grammar leaves it outside the command node. Display math (\[...\]) bridges into surrounding prose as an exclusion zone.

• R Sweave — Preprocesses R code chunks (<<...>>= through @) by blanking them with whitespace, then delegates to the LaTeX extractor.

• reStructuredText — Extracts paragraph and title nodes, with inline literals as exclusion zones. A directive is split by role: its argument is a path or language (except on csv-table, admonition and friends, where it is the rendered title), its :caption: and :alt: options are prose, and its body is prose unless the type is code or data (code-block, math, raw, toctree, …). Because tree-sitter-rst leaves a directive body unparsed, the body is recovered from the raw lines: an explicit markup line, a doctest >>> prompt or a trailing :: takes its whole indented block out of the prose, so a code block nested in a .. note:: stays out.

• Org mode — Extracts paragraph text and heading titles, plus the parts of a structured node that render: the contents of a #+begin_quote or #+begin_verse block, table cells, footnote definitions ([fn:1] …) and the value of a #+TITLE:, #+SUBTITLE:, #+CAPTION: or #+DESCRIPTION: directive. Skips #+begin_src blocks, drawers (:PROPERTIES:), LaTeX environments, comments, and the remaining directives.

• BibTeX — Extracts prose from specific fields: title, booktitle, abstract, note, annote, annotation, howpublished, and series. Other fields (author, journal, year, etc.) are ignored. LaTeX commands inside values (e.g. \emph{...}) are handled via exclusion zones.

• Typst — Collects text nodes from paragraphs, headings, and list items. Skips code blocks (```), inline code (`), math ($...$ and $ ... $), #code expressions, set/show rules, let bindings, imports, includes, labels, references, URLs, escapes, and comments. Content blocks passed to a function (#columns(2)[...], #align(center)[...], #figure(caption: [...])[...]) are still checked — only the call itself and its non-content arguments are skipped. Inline markup (*bold*, _italic_) is bridged through.

• Forester — Collects text and escape nodes, skipping math (#{...}, ##{...}), verbatim fences, wiki links, comments, and structural commands (\import, \ref, \def, etc.). Display math bridges as an exclusion zone.

Adding more file types

You can add support for extra file types without code in two ways:

• map new extensions onto existing built-in language IDs, or

• define regex-based Simplified Language Schema (SLS) YAML files in .langcheck/schemas/.

See the Config-Only Language Guide for both workflows, including a full schema example.

Tip

To add support for an entirely new markup language with its own tree-sitter grammar, see the Plugin Language Guide.

Checking Languages

The spell-check and grammar-check language is separate from the file type. Click the language indicator in the VS Code status bar to switch:

• EN-US — American English

• EN-GB — British English

• DE-DE — German

• FR — French

• ES — Spanish

Language detection can also be automatic via the whatlang crate when no explicit language is set.

More than one language in one file

engines.spell_language is the document default, not the only answer. A thesis in French quoting English, or a German paper with an English abstract, is checked passage by passage: the prose is grouped by language and each group goes to the engines that read it, so Harper handles the English while LanguageTool handles the French.

Three things declare a language, strongest first.

A region directive. Works in every format, and is the only option where the markup has nothing to say:

<!-- lang-check-begin lang:fr -->
Ceci est du texte français.
<!-- lang-check-end -->

A scope marker, which runs until the next one:

<!-- lang: fr -->

The markup’s own declaration, where the format has one. Typst does:

#set text(lang: "fr", region: "CH")   // the rest of the enclosing block
#text(lang: "en")[An English aside.]  // just this content

That last one needs no extra annotation at all — the lang: an author already writes for hyphenation and quotation marks is the one the checker reads, and region: becomes the BCP-47 subtag.

See the directives reference for the full syntax, and examples/typst/ for a worked document.

When nothing can read the language

Not every language has an engine. LanguageTool has no Hebrew and no Latin, and Harper reads only English. Rather than let the passage pass as clean, the checker reports a languagecheck.no-provider diagnostic naming the language, and does not mark the engine unhealthy for declining one.

For most of those, Hunspell fills the gap with spelling — Hebrew, Latin, Welsh, Old English and a long tail besides, using the dictionary format LibreOffice and Firefox already read. It checks spelling and nothing else, so where LanguageTool supports a language, prefer it.

In VS Code the editor offers to install a dictionary when it meets a language none of the enabled engines read, and remembers a refusal. An engine that declares the language answers for it, so a custom engines.external checker naming languages: ["cy"] covers Welsh and the offer never appears — but one that declares nothing at all still claims every language, and suppresses the offer for all of them. See External providers for what a provider should declare.

Bare tags

A declaration with no region is resolved before it reaches the engines: it takes the document default’s region when the language agrees, so lang: "en" under an en-GB document means en-GB, and a known variant otherwise. The default is resolved the same way, so spell_language: en checks as en-US and spell_language: de as de-DE. In VS Code the resolved region is shown as ghost text after the tag, so lang: "en" reads en-US where it is written; the inlay hint toggle hides it along with the suggestion hints.

This is not cosmetic. LanguageTool accepts bare en and de and then reports no spelling errors at all in them, so an unresolved tag would check nothing and look clean.