# Latgalian MCP server

> A read-only MCP server and JSON API over the English Wiktionary's Latgalian
> entries — Latgalian is a living East Baltic language of Latgale, in eastern
> Latvia (ISO 639-3 ltg; Latin script with diacritics). Use it to read and write
> Latgalian with attested words and correct inflected forms instead of inventing
> them. Read-only.

## What it does
- Looks a Latgalian word up (any form) and returns its senses plus full
  declension/conjugation tables — the seven noun cases, the adjectives'
  indefinite/definite × gender family, and the verbs' positive and negative
  conjugations.
- Finds the Latgalian equivalents of an ENGLISH word (search_language=eng): each
  sense's translations from Wiktionary's translation tables (the result's
  translations block), expanded into their full dictionary entries.
- Resolves an inflected query to its dictionary lemma automatically via paradigms
  already cached — the result says so (search_method=lemma_index, with
  resolved_from/resolved_lemma). No external analyzer exists for Latgalian, so a
  form not yet in any cached paradigm won't resolve — look the lemma up directly.
- Folds diacritics, so ASCII spellings work once a paradigm is cached: every
  Latgalian diacritic strips to its base letter ('akmins' finds 'akmiņs'). This
  is a one-way fold — diacritics optional, never a two-way digraph swap.

## How to connect (MCP)
Remote streamable-HTTP endpoint (no auth, read-only): https://latgalian.valksor.com/mcp
  Claude Code:  claude mcp add --transport http latgalian https://latgalian.valksor.com/mcp
  Cursor / Cline (~/.cursor/mcp.json):
    {"mcpServers": {"latgalian": {"url": "https://latgalian.valksor.com/mcp"}}}
  stdio-only clients — bridge with mcp-remote:
    {"mcpServers": {"latgalian": {"command": "npx",
       "args": ["-y", "mcp-remote", "https://latgalian.valksor.com/mcp"]}}}
Tools (each returns Markdown plus typed structuredContent per its outputSchema):
- latgalian_search(query, search_language='ltg'|'eng', include_forms=true, max_forms=N)
    Senses plus inflection handles (entry_id = the lemma/page title, word_class =
    the part-of-speech section). With include_forms (default TRUE on MCP; the JSON
    API defaults it false), each match's full table comes back inline. With
    search_language='eng' the query is an English word and the result lists its
    per-sense Latgalian equivalents plus their entries (English→Latgalian).
- latgalian_get_inflections(entry_id, word_class)
    The full paradigm for one handle. Its structuredContent nests the paradigm
    under a "result" key (the search tool's payload is unwrapped); the JSON API
    returns it directly. Both shapes are discriminated by "category":
    nominal gives case keys -> {singular, plural} (nouns use the bare case;
    adjectives prefix the state and suffix the gender — 'indefinite nominative
    masculine'); verbal gives finite "<mood> [tense]" -> "<person> <number>" ->
    form (the negative table's keys carry a 'negative ' prefix) plus the
    non-finite forms; "not_found" when no table exists.
- latgalian_romanize(text)
    Pure local diacritic→ASCII fold for an English-trained downstream (a voice/TTS,
    a search box) — no network, no cache, output always ASCII. Lowercases and maps
    each special letter (ā→aa, š→sh, ž→zh, ļ→ly, ņ→ny); diacritics are dropped,
    never added.
Search tips: one Latgalian word per call; any form works, and ASCII spellings
('akmins') resolve once the lemma's paradigm is cached; for English→Latgalian
pass search_language='eng' (multiword entries like 'apple tree' work).
Prompts & resources (MCP):
- Prompts latgalian_translate_to(text) and latgalian_translate_from(text, target_language)
  package the grounded-translation method.
- Resource latgalian://glossary/grammar-labels documents the paradigm keys; this
  document is the llms.txt resource.

## How to call (JSON API)
- GET https://latgalian.valksor.com/api/search?query=akmi%C5%86s
- GET https://latgalian.valksor.com/api/search?query=stone&search_language=eng
- GET https://latgalian.valksor.com/api/inflections?entry_id=byut&word_class=verb
- GET https://latgalian.valksor.com/api/romanize?text=akmi%C5%86s
- GET https://latgalian.valksor.com/api/health
- Interactive docs: https://latgalian.valksor.com/api/docs  ·  schema: https://latgalian.valksor.com/api/openapi.json

## Caching and politeness
Dictionary responses are cached server-side with a long TTL (Wiktionary content
changes slowly). Upstream calls are serialised with a polite minimum interval.
Calls are rate limited per client IP — on a 429, wait a few seconds and retry.
The romanize endpoint is a pure local transform and is exempt from the rate limit.

## Source and licence
- Dictionary: Wiktionary, the free dictionary (en.wiktionary.org), Latgalian entries. Wikimedia Foundation. Text available under CC BY-SA 4.0; attribution and share-alike required when republishing.
- Licence: CC BY-SA 4.0 — attribute and share alike when republishing dictionary text.