Gujarati Lexicon Mcp

Gujarati Lexicon MCP Server [PyPI]

A grounded Gujarati dictionary for AI assistants. Connect it to Claude (or any MCP client) and the assistant looks Gujarati words up in real dictionary data instead of guessing: meanings, synonyms, idioms (રૂઢિપ્રયોગ), and inflected forms like ઘરમાં or આંખોમાં.

Why this exists: LLMs are unreliable with Gujarati. They invent meanings, mix up idioms, and produce convincing but wrong usage. This server gives the model facts to stand on and tells it plainly when a word is not in the dictionary.

[Demo]

Example

User: અમે નો અર્થ શું છે?

Claude (after calling define): અમે means "we", the plural of હું (I)…

User: ઘરમાં નો અર્થ?

Claude: ઘરમાં = ઘર + માં → "in the house". The server matched the inflected form to its base word ઘર.

Features

• About 6,800 words with English meanings, from Wiktionary
• Hand-checked entries with Gujarati meanings and idioms (curated set, growing)
• Inflection handling: ઘરમાં, ઘરે, આંખોમાં, and અમે resolve to their base words
• "Did you mean?" suggestions for typos (પાણિ → પાણી)
• Source attribution in every response, so the assistant can say where a meaning came from

Quick start

Requires uv.

Add this to your Claude Desktop config (Settings → Developer → Edit Config):

      "args": ["gujarati-lexicon-mcp"]

Restart Claude Desktop completely (quit from the system tray), open a new chat, and ask about any Gujarati word.

On Windows, if Claude Desktop can't find uvx, use its full path (find it with where uvx), e.g. C:\\Users\\<you>\\.local\\bin\\uvx.exe.

How lookup works

A word goes through three layers, from most to least reliable:

flowchart LR    A[Input word] --> B{Exact headword?}    B -- yes --> C{Pure inflected form?}    C -- yes --> D[Follow Wiktionary link<br/>ઘરે → ઘર]    C -- no --> E[Return entry]    D --> E    B -- no --> F[Strip suffixes<br/>માં, નો, ની, ે, ો ...]    F -- match --> E    F -- no match --> G[Not found + suggestions<br/>tell the model not to guess]

1. Exact match. A real dictionary entry always wins over a guessed base form.
2. Wiktionary inflection links. Wiktionary records that ઘરે is the locative of ઘર. The converter reads the structured form_of field, and falls back to parsing glosses like "plural of X" where that field is missing.
3. Suffix stripping. As a last resort, common endings are removed (up to two, e.g. આંખોમાં → આંખો → આંખ). A stripped form is accepted only if it is a real headword, so words are never mangled.

Responses keep sources separate (curated vs wiktionary_senses), so the model can attribute meanings honestly.

Data sources

Wiktionary data comes from Wiktionary via kaikki.org, extracted with wiktextract:

Tatu Ylonen. Wiktextract: Wiktionary as Machine-Readable Structured Data. Proceedings of the 13th Conference on Language Resources and Evaluation (LREC), pp. 1317–1325, 2022.

Rebuilding the Wiktionary data

The raw dump is not committed. To regenerate wiktionary.json:

mkdir -p data/rawcurl -L -o data/raw/kaikki-gujarati.jsonl "https://kaikki.org/dictionary/Gujarati/kaikki.org-dictionary-Gujarati.jsonl"uv run python scripts/build_wiktionary.py

The converter groups entries by word, keeps only what the tools need (20 MB → a compact JSON file), drops non-Gujarati-script headwords, and records inflection links.

Development

git clone https://github.com/aarshbharat/gujarati-lexicon-mcpcd gujarati-lexicon-mcpuv syncuv run pytest -v                                    # run testsuv run mcp dev src/gujarati_lexicon_mcp/server.py   # open MCP Inspector

Project layout:

src/gujarati_lexicon_mcp/├── server.py           # MCP server: tools, resource, prompt└── data/               # dictionary data shipped with the packagescripts/build_wiktionary.py   # raw kaikki dump → wiktionary.jsontests/                        # pytest

Known limitations

• Verb forms are only partly covered: forms Wiktionary links (e.g. હોઈશ → હોવું) work, but phrases like લે છે are not lemmatized.
• Wiktionary synonyms are merged across senses. ઘર's list includes ઓફિસ (from its "office" sense). The tool tells the model this.
• The curated set is small. Most entries have English meanings only; Gujarati-language meanings and idioms come from the hand-checked data.
• Grounding covers facts, not everything the model says. The model may add correct background from its own knowledge (e.g. the inclusive/exclusive "we" distinction for અમે/આપણે). The server's instructions ask it to label general knowledge, but cannot force it.

[PyPI]

Roadmap

• Grow the curated set, especially idioms and proverbs (કહેવત)
• Sense-level synonyms instead of a merged list
• Better verb lemmatization
• Streamable HTTP transport and a hosted endpoint
• Publish to PyPI and the MCP Registry

License

• Code: MIT
• Wiktionary-derived data (src/gujarati_lexicon_mcp/data/wiktionary.json): CC BY-SA, see Wiktionary:Copyrights

Built by Aarsh Dhokai · AI × AI: Artificial Intelligence × Aarsh India