cymraeg-mcp 0.1.0

Welsh proofing/validation MCP server — deterministic spelling (hunspell cy_GB) and mutation (treiglad) checking, 100% offline

cymraeg-mcp

Gweinydd MCP prawfddarllen Cymraeg / Welsh proofing & validation MCP server

English below

---

Cymraeg

Haen ddilysu benderfynadwy ar gyfer testun Cymraeg: gwirio sillafu (hunspell cy_GB) a gwirio treigladau (peiriant rheolau lleol). 100% lleol — dim APIau allanol, dim allweddi API, dim lawrlwytho wrth redeg, yn gweithio'n gyfan gwbl all-lein.

Nid pecyn cymorth Cymraeg cyffredinol mo hwn. Mae LLMau eisoes yn ysgrifennu Cymraeg dda — y broblem yw nad oes modd gwarantu cywirdeb allbwn tebygolaethol. Yr egwyddor yma yw: "Claude sy'n drafftio, hwn sy'n dilysu." Wedi'i anelu at unrhyw un sy'n cyhoeddi cynnwys dwyieithog dan Safonau'r Gymraeg — cynghorau, GIG Cymru, cyrff cyhoeddus — sydd angen gramadeg gwarantedig-gywir.

Offer

Offeryn	Mewnbwn	Allbwn
cymraeg_check_spelling	testun Cymraeg	camsillafiadau + awgrymiadau (hunspell cy_GB Prifysgol Bangor)
cymraeg_check_mutations	testun Cymraeg	gwallau treiglo (meddal / trwynol / llaes) gyda chywiriadau, hyder, a'r rheol yn ddwyieithog
cymraeg_explain_mutation	gair + cyd-destun	pa dreiglad sy'n berthnasol a pham, ynghyd â thabl treiglo llawn y gair
cymraeg_proof_text	testun Cymraeg	adroddiad cyfun: sillafu + treigladau mewn un pas

Gosod

Gweler y cyfarwyddiadau gosod yn yr adran Saesneg isod — mae'r gorchmynion yn union yr un fath.

Cyfrannu

Mae'r rheolau treiglo yn data/mutation_rules.yaml — ychwanegwch sbardunau, enwau benywaidd, neu eithriadau yno a chyflwynwch PR. Rhedwch uv run pytest cyn cyflwyno.

---

English

A deterministic validation layer for Welsh text: spellchecking (hunspell cy_GB) and mutation checking (a local rules engine). 100% local — no external APIs, no API keys, no first-run downloads, fully offline.

This is NOT a general Welsh toolkit. LLMs already write good Welsh — the problem is that probabilistic output can't be guaranteed correct. The positioning here is: "Claude drafts it, this verifies it." Target users: anyone publishing bilingual content under the Welsh Language Standards (councils, NHS Wales, public bodies) who need guaranteed-correct grammar rather than probabilistic LLM output.

Novelty: as of July 2026, no other Welsh-language MCP server exists.

Tools

Tool	Input	Output
cymraeg_check_spelling	Welsh text	misspellings + up to 5 suggestions each (Bangor University hunspell cy_GB, 72k+ entries, bundled)
cymraeg_check_mutations	Welsh text	mutation errors (treiglad meddal / trwynol / llaes) with corrections, confidence levels, and the rule cited bilingually
cymraeg_explain_mutation	word + context	which mutation applies and why, whether the word is radical or mutated, and the word's full mutation table
cymraeg_proof_text	Welsh text	combined report: spelling + mutation issues in one pass

All tools are read-only, idempotent, and fully offline.

What the mutation engine covers

• Soft mutation (treiglad meddal): after the prepositions am, ar, at, dan, tan, dros, drwy/trwy, gan, heb, hyd, i, o, wrth; after dy, dau, dwy, neu, mor, rhy, pur, go, dyma, dyna, dacw, pa; feminine singular nouns after y/yr/'r/un (using an extensible noun list); with the ll/rh exceptions applied.
• Nasal mutation (treiglad trwynol): after fy; yn/ym/yng locative with form agreement (flags "yn Nghaerdydd" → "yng Nghaerdydd", "ym Bangor" → "ym Mangor"); blynedd/blwydd after numerals (pum mlynedd).
• Aspirate mutation (treiglad llaes): after â, gyda, tua, tri, chwe, a (and), ei (her), na.
• Ambiguity handled honestly: "ei" (his/her) is disambiguated by an echoing pronoun where present ("ei tad hi" → "ei thad hi"); "a" offers both the aspirate (conjunction) and soft (relative pronoun) readings; genuinely ambiguous cases are reported at low confidence and hidden by default.

The engine is tuned for precision over recall: it only flags what it can justify from an explicit trigger rule plus a dictionary check, and every finding carries a confidence level. A clean result does not guarantee every mutation is correct — see limitations.

Install

Requires Python 3.11+. With uv (preferred):

git clone https://github.com/dwain-barnes/cymraeg-mcp
cd cymraeg-mcp
uv sync

Or with pip:

pip install .

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "cymraeg": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/cymraeg-mcp", "cymraeg-mcp"]
    }
  }
}

Claude Code

claude mcp add cymraeg -- uv run --directory /path/to/cymraeg-mcp cymraeg-mcp

Example

Input: cymraeg_proof_text("Dw i eisiau mynd i Caerdydd gyda fy cath.")

Output: 2 mutation issues — Caerdydd → Gaerdydd (soft mutation after the preposition 'i'), cath → nghath (nasal mutation after 'fy').

Limitations (read this)

• A clean report is not a proof of correctness. The engine deliberately stays silent where Welsh is genuinely ambiguous without deeper parsing — e.g. unmutated words after predicative/locative "yn", gender of nouns not in its list, mutation after inflected verbs (object soft mutation), and wrong-mutation-type cases like "gyda dad".
• The feminine-noun rule only covers nouns listed in data/mutation_rules.yaml (extensible by PR).
• The dictionary includes correctly mutated forms, so the spellchecker will not flag a mutation error as a misspelling — that's the mutation checker's job.
• hunspell cy_GB is comprehensive (72k+ stems plus affix rules) but no dictionary is complete; rare proper nouns may be reported as unknown.

Licensing

• Code: MIT © Dwain Barnes (EryriLabs).
• Bundled dictionary (data/dictionaries/cy_GB.*): © Bangor University (Prifysgol Bangor) and Troi, licensed under the LGPL v3 or later — see data/dictionaries/LICENCE. It is redistributed unmodified with its licence and attribution intact, as permitted by the LGPL.
• No data leaves your machine; there is no telemetry and no network access.

Contributing

Mutation rules live in data/mutation_rules.yaml — triggers, feminine nouns, exceptions and never-mutate words are all data, not code. Add a rule, add a test case to tests/test_mutations.py, and run:

uv run pytest
uv run ruff check src tests

Credits

Built by Dwain Barnes / EryriLabs — HuggingFace. Dictionary by Uned Technolegau Iaith, Prifysgol Bangor.