Skip to main content
Glama
aynumosir

ainu-mcp

by aynumosir
OverviewSchemaRelated ServersScoreDiscussions
TypeScript
Remote

ainu-mcp

A hosted Model Context Protocol server for the Ainu-language toolchain (aynumosir/ainu-mcp). It lets an LLM (Claude Code, Claude Desktop, etc.):

  • Edit the Itak-uoeroskip glossary directly in its Google Sheets source of truth — search, add, and update entries from chat.

  • Reference all the other Ainu materials in one place — search ~195 k aligned corpus sentences, look up words across 11 dictionaries, scan the grammar bibliography, and convert between Latin/Katakana/Cyrillic scripts.

  • Research an entry in one call — entry_research(word) composes all of the above into a single structured response, so the model can draft a well-grounded glossary entry without round-tripping.

How this repo is laid out

There is one MCP server — the hosted Cloudflare Worker — and a Python ETL that feeds it. (An earlier local stdio server was retired; the Worker is now the single MCP surface.)

Part

What it is

worker/

The MCP server: a Cloudflare Worker (Streamable-HTTP MCP) at mcp.aynu.org, GitHub-OAuth gated, reading a Turso (libSQL, FTS5 trigram) reference store. See worker/README.md for the deploy guide.

src/ainu_mcp/ + etl/build_d1.py

The Python ETL: corpus / dictionary / grammar / glossary loaders that bake the reference data into the Worker's Turso DB. Not a server.

Access model: any GitHub user who authenticates gets the read/reference tools; members of the aynumosir org (or anyone in ALLOWED_USERS) additionally get the glossary write + maintenance tools. Non-members never see the write tools at all.

Related MCP server: Japanese Vocab Anki MCP Server

Connecting

The server speaks Streamable-HTTP MCP and authenticates over GitHub OAuth (your client opens a browser the first time).

Project-scoped (.mcp.json in any project where you want it available — this is what this repo ships):

{
  "mcpServers": {
    "ainu": {
      "type": "http",
      "url": "https://mcp.aynu.org/mcp"
    }
  }
}

Or user-scoped — add the same block to ~/.claude.json under mcpServers.

Building / refreshing the reference data

The corpus, dictionaries, and grammar tables in Turso are built by the Python ETL, which the Worker only reads. To rebuild the seed:

Requires Python ≥ 3.13 and uv.

uv sync
AINU_ROOT=/home/mkpoli/projects/Ainu uv run python etl/build_d1.py

AINU_ROOT (default /home/mkpoli/projects/Ainu) must contain:

  • ainu-corpora/data.jsonl

  • ainu-dictionaries/<dict-name>/*.tsv

  • ainu-grammar/{books,articles}/...

(The stopword list from aynumosir/ainu-stopwords is public, so the ETL fetches it from GitHub automatically — no checkout needed. A local ainu-stopwords/ainu-stopwords.txt under AINU_ROOT, if present, is used instead. The morphology tools do not feed into this seed — they proxy the MDB forms engine live, see the Morphology section below.)

The seed is loaded into Turso with the batched libSQL loader (see docs/REFRESHING-DATA.md). A scheduled GitHub Action (refresh-reference-data.yml) rebuilds and reseeds Turso monthly. The live glossary is read straight from Google Sheets, so glossary edits do not depend on this refresh.

Tool surface

Glossary (read + write — the editing loop)

Tool

Purpose

glossary_list_categories

List sheet tabs with description and entry count

glossary_list_entries(category, limit, offset)

Page through entries in a category

glossary_get_entry(category, row)

Read one entry (returns row_hash for safe editing)

glossary_search(query, fields?, category?, limit?)

Substring search; optionally scoped to columns / one category

glossary_add_entry(category, fields)

Append a new row

glossary_update_entry(category, row, fields, expected_row_hash?)

Update cells, with optimistic locking

glossary_untranslated(category?, langs?, limit?)

Find rows missing 日本語/English/中文 — translation worklist

glossary_audit()

Find inconsistencies: =an + N1, parens, transitivity mismatch, duplicates, …

glossary_missing_high_frequency(top_n?, min_count?)

Frequent corpus tokens that are dictionary-attested but not in glossary — vocab-gap worklist

glossary_refresh_site_cache(dry_run?)

Republish table.json/sheets.json to Cloudflare R2 so itak.aynu.org reflects edits immediately (instead of waiting for the weekly cron)

Write + maintenance tools (glossary_add_entry, glossary_update_entry, glossary_audit, glossary_missing_high_frequency, glossary_refresh_site_cache) are only available to aynumosir org members.

Optimistic locking: pass the row_hash you got from glossary_get_entry or glossary_search as expected_row_hash when updating. If someone else edited the row since, the update is refused — re-read and retry.

Reference (corpus, dictionaries, grammar)

Tool

Purpose

corpus_search(query, lang, dialect?, author?, limit?)

Search aligned Ainu/Japanese sentences (lang: ain, jpn, any)

corpus_stats

Total sentences + top dialect distribution

corpus_word_frequency(word)

Corpus frequency of a word — count, rank, stopword flag, corpus totals (affix-clitics normalized, so ku=nukar = nukar)

corpus_frequency_list(limit?, offset?, include_stopwords?, min_count?)

Ranked frequency list; set include_stopwords=false to list content words only

corpus_stopwords

The Ainu stopword list (from aynumosir/ainu-stopwords)

dictionary_list

List dictionaries with entry counts

dictionary_lookup(word, dicts?, fields?, limit?)

Multi-dictionary lookup (any field; supports substring)

dictionary_reverse_lookup(aynu, dicts?, limit?)

Ainu → Japanese/English by exact lemma first then substring; Ota's reverse index included

grammar_list(kind?)

List grammar books / articles

grammar_search(query, include_transcribed?, limit?)

Filename/title/author search + fulltext over transcribed sources

Morphology (possessed / plural / derived forms)

All three morphology tools are thin proxies to the Ainu Morpheme Database forms engine (mdb.aynu.org/api/forms, over the env.MDB service binding). The generative engine — possessed-noun forms, plural verb forms, and derivations — and its data live in MDB; this server holds no morphology copy of its own. The engine is hybrid + provenanced: rules generate, harvest + curated exceptions validate/override, and every form is tagged source (rule | attested | exception) + confidence. A source='rule' form with no attested_ref is predicted-but-unattested — surfaced as a discovery aid but flagged.

Tool

Purpose

morphology_search(query, category?, limit?)

Search forms — query matches the surface form, its analysis/decomposition, or the lemma (substring, via /api/forms?q=); filter by category (possessed/plural/derived, mapped to the upstream relation facet)

morphology_reverse_lookup(base, category?, limit?)

From a base lemma to the forms built on it (e.g. sapa → sapaha); base is an exact lemma_id match (via /api/forms?lemma=)

morphology_forms(lemma, category?, relation?, feature?, provenance?, limit?)

Look up a lemma's possessed-noun forms (sapa → sapaha), plural verb forms (-pa/suppletive, role-sensitive object vs subject number) and derivations. Filter by category (domain: nominal/verbal), relation (possessed/plural/derived), feature (a feature-bundle facet), provenance

Localization (i18n strings)

Tool

Purpose

localizations_search(query, project?, lang?, limit?)

Search how real Ainu-language software has translated its UI — query matches the Ainu text, the source-language original, or the message key; filter by project slug or lang (ain, ain-Latn, …)

localizations_list_projects

List the indexed projects (repo, format, source language, string count)

Aynuwiki (Ainu encyclopedias)

Live access to the two Ainu-language wikis — Aynuwiki and the Ainu Wikipedia in the Wikimedia Incubator (Wp/ain/) — over their MediaWiki APIs (no snapshot; always current).

Tool

Purpose

wiki_search(query, site?, limit?)

Search articles; site: aynuwiki, incubator, or both (default)

wiki_get_page(title, site?)

Full article text (clean plain-text extract, or raw wikitext for template-heavy pages)

Script conversion

Tool

Purpose

convert_script(text, from, to)

Convert between latn / kana / cyrl

detect_script(text)

Detect script of a string

script_all(text)

Return all three script renditions in one call

Composed

Tool

Purpose

entry_research(word, ...)

One-shot: scripts + syllables + glossary hits + dictionary hits + corpus examples

Example session

Typical edit flow Claude would run:

  1. entry_research("kunne") → see existing dictionary defs, corpus contexts, current glossary entries

  2. glossary_search("kunne") → confirm the row(s) and grab row_hash

  3. glossary_update_entry("色", 47, {"English": "black"}, expected_row_hash="…")

For the full editing convention (transitivity, gloss style, sources), see AGENTS.md.

This server cannot be installed

F
license - not found
-
quality - not tested
B
maintenance

How are these scores calculated?

Maintenance

–Maintainers
–Response time
–Release cycle
–Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/aynumosir/ainu-mcp'

If you have feedback or need assistance with the MCP directory API, please join our Discord server