# Livonian MCP server > A read-only MCP server and JSON API over the real Livonian–Estonian–Latvian dictionary at > livonian.tech. Use it to translate to/from Livonian with ATTESTED words and > grammatical forms instead of inventing them. Read-only. ## What it does - Looks up words in the Livonian dictionary (query in English, Livonian, Latvian or Estonian) and returns attested headwords, parts of speech, English glosses and example sentences. - Returns full declension/conjugation (inflection) tables for a given word. - Resolves an inflected Livonian query (search_language=liv) to its dictionary lemma when the form is known — the upstream search matches headwords only, so e.g. "kastāb" alone finds nothing, but it resolves to "kastõ" and returns that entry (marked search_method=lemma_index, with resolved_from/resolved_lemma). ## How to connect (MCP) Remote streamable-HTTP endpoint (no auth, read-only): https://livonian.valksor.tools/mcp Native remote-url clients: Claude Code: claude mcp add --transport http livonian https://livonian.valksor.tools/mcp Cursor / Cline (~/.cursor/mcp.json): {"mcpServers": {"livonian": {"url": "https://livonian.valksor.tools/mcp"}}} VS Code (.vscode/mcp.json): {"servers": {"livonian": {"type": "http", "url": "https://livonian.valksor.tools/mcp"}}} stdio-only clients (Claude Desktop, Windsurf, Codex) — bridge with mcp-remote: {"mcpServers": {"livonian": {"command": "npx", "args": ["-y", "mcp-remote", "https://livonian.valksor.tools/mcp"]}}} (Claude Desktop rejects a bare "url"; Pro+ can instead add it under Settings -> Connectors -> Add custom connector and paste https://livonian.valksor.tools/mcp.) Tools: - livonian_search(query, search_language=eng|liv|lat|est, include_forms=true, max_forms=N) Returns dictionary entries plus "inflection handles" (entry_id + word_class). With include_forms (default TRUE on the MCP surface; the JSON API defaults it false), each match's full inflection table comes back inline, so one call usually suffices. - livonian_get_inflections(entry_id, word_class) word_class is one of s (noun), adj, vi (intransitive verb), vt (transitive verb); any other value gets a 422. entry_id here is a NUMERIC id (e.g. 12127), not the headword string — unlike the sibling Wiktionary dictionaries; it only ever comes from a search result's handle. Returns the full inflection paradigm for a handle from a search result. The MCP surface renders Markdown and carries the same structure as structuredContent under a "result" key (per the tool's outputSchema; the other tools' payloads are unwrapped); the JSON API returns it directly. Both shapes are discriminated by "category": a nominal (noun/adjective/numeral/pronoun) gives cases -> {singular, plural}; a verb gives finite moods -> person -> {affirmative, negative} (the negation auxiliary folded into the negative) plus non-finite sections -> label -> form; category "not_found" when the word has no paradigm. - livonian_romanize(text) Strips Livonian orthography (macrons, diacritics, capitals) to clean, pronounceable ASCII. Palatalized consonants (ņ ļ ŗ ţ ḑ) soften before a vowel (ņ+a -> nya) but go plain elsewhere (ikšiggiņ -> ikshiggin); the broken tone (si'zzõl -> siz-zul) and long-vowel diphthong glides (kīen -> kee-en) become hyphens. For an English-trained downstream (a voice/TTS, a search box). Pure transform, no dictionary lookup. - livonian_deromanize(text) The inverse direction: given clean-ASCII romanized text, find the real Livonian word(s) whose romanization matches. romanize is lossy, so this is a LOOKUP against words already cached — attested spellings come back as source=dictionary. On a miss, one best-effort reconstruction (source=reconstruction, approximate=true), a guess. Coverage grows as words are cached. Typical flow (English -> Livonian): one livonian_search("wet") call returns entries AND each match's inflection table inline, so it usually suffices; call livonian_get_inflections only for a match a broad search left un-expanded (expansion is bounded by a cap — raise max_forms). Search tips: query one content word, not a phrase — and only one per call. Even with several candidate words or synonyms, search them one at a time ("see", then "look"), never combined into one query like "see look". A multi-word query often returns nothing even when the word exists. Search "water", not "to water"; drop "to"/"the"/"a" and infinitive markers, then refine with the main word. Prompts & resources (MCP): - Prompts livonian_translate_to(text) and livonian_translate_from(text, target_language) package the grounded-translation method above. - Resource livonian://glossary/word-classes explains the word_class codes on search handles; this document is the llms.txt resource. ## How to call (JSON API) - GET https://livonian.valksor.tools/api/search?query=wet&search_language=eng - GET https://livonian.valksor.tools/api/inflections?entry_id=12127&word_class=s - GET https://livonian.valksor.tools/api/romanize?text=L%C4%ABv%C3%B5%20k%C4%93%C4%BC (URL-encoded "Līvõ kēļ") - GET https://livonian.valksor.tools/api/deromanize?text=kaal (attested spellings; reconstruction on a miss) - GET https://livonian.valksor.tools/api/health - Interactive docs: https://livonian.valksor.tools/api/docs · schema: https://livonian.valksor.tools/api/openapi.json ## Languages search_language selects the language your QUERY is in: eng (default), liv, lat, est. Headwords are always Livonian; glosses are English. ## Limits & caching - Rate limited per client IP — do not use this as a bulk scraping proxy. The JSON /api/romanize endpoint is exempt (a free local transform); on the MCP surface all tools share the one per-IP bucket, so livonian_romanize still counts against it. - Results are cached permanently (the dictionary is static); repeat lookups are served from cache and never re-fetch upstream. ## Attribution (required whenever you share or redistribute this material) Data from the Livonian.tech platform (University of Latvia Livonian Institute), licensed CC BY-SA 4.0; adaptations must be shared under the same licence. Cite the specific database you used (per https://livonian.tech/en/referencing-and-quoting/): Word lookups — Livonian lexicographic database: Ernštreits, Valts (ed. in chief), Viitso, Tiit-Rein, Vāvere, Signis, Damberg, Pētõr, Kurpniece, Milda, Balodis, Uldis, Tuisk, Tuuli, Kūla, Gita, Tomingas, Marili, Soosaar Sven-Erik, Sedláčková, Anna & Jurgenovskis, Toms. 2024. Livonian lexicographic database. Riga: University of Latvia Livonian Institute. https://livonian.tech/ Inflection tables — Livonian morphology database: Ernštreits, Valts (ed. in chief), Viitso, Tiit-Rein, Kurpniece, Milda & Vāvere, Signis. 2024. Livonian morphology database. Riga: University of Latvia Livonian Institute. https://livonian.tech/