The multilingual-dictionary-mcp is an offline-first MCP server providing 28 dictionary and linguistic tools across thousands of languages. It downloads ~15.6 GB of data once and operates with no internet, rate limits, or outages thereafter.
Data Management
dictionary_status: Check the offline data installation state
dictionary_install: Manually trigger or re-trigger the data download
Definitions, Etymology & Pronunciation (Wiktextract — 4,755 languages, 10.5M entries)
dictionary_lookup: Full definitions, part-of-speech, IPA, etymology, and translations
dictionary_summary: Concise plain-text summary of a word's senses
dictionary_etymology / dictionary_pronunciation: Extract etymology text or IPA pronunciations
dictionary_search: Prefix-search the corpus when spelling is uncertain
dictionary_random: Fetch a random word entry, optionally filtered by language
Word Relations (ConceptNet — 80+ languages)
Synonyms, antonyms, hypernyms (broader concepts), hyponyms (narrower concepts), meronyms (parts), holonyms (wholes)
Derivation, etymological cognates, typical uses, typical actions, and typical locations
dictionary_all_relations: Retrieve all relation types in a single call
dictionary_translate: Translate words via cross-lingual synonym links
Embedding-Based Semantic Search (Numberbatch — 78 languages)
dictionary_related: Find semantic neighbors via cosine similarity
dictionary_semantic_neighbors: Explicit embedding lookup with optional cross-lingual filtering
dictionary_means_like: Find words/phrases with approximately the same meaning across 78 languages
English-Specific Tools (CMU Pronouncing Dictionary)
dictionary_rhymes: Find perfect or near rhymes
dictionary_sounds_like: Discover homophones and soundalikes
dictionary_spelled_like: Match words by spelling pattern using ? and * wildcards (useful for crosswords)
dictionary_suggest: Autocomplete suggestions for partial words
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@multilingual-dictionary-mcp synonyms of 'amor' in Spanish"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An offline-first MCP server for multilingual dictionary lookups — definitions, synonyms, antonyms, hypernyms, hyponyms, meronyms, translations, etymology, pronunciation, semantic neighbors, rhymes, and more — across 4,755 languages.
No third-party APIs. No rate limits. No outages. The server downloads its data from a CDN once on first run and never goes online again.
Source | Size | What it covers |
Wiktextract (Kaikki.org) | 6.6 GB | 10.5M dictionary entries across 4,755 languages — definitions, etymology, IPA, translations |
ConceptNet 5.7 | 5.6 GB | 24.3M semantic edges across 80+ languages — Synonym, Antonym, IsA, PartOf, RelatedTo, UsedFor, CapableOf, AtLocation, DerivedFrom, EtymologicallyRelatedTo |
Numberbatch embeddings | 3 GB | 9.16M concepts × 300d multilingual embeddings — semantic neighbors via cosine similarity |
CMU Pronouncing Dictionary | 19 MB | English rhymes, soundalikes, autocomplete, spell patterns |
Total bundle: ~21 GB (medium profile, default).
Related MCP server: word-orb
The whole bundle ships in 6 artifacts. Total 5.4 GB on the wire (gzip-compressed where it helps), 15.6 GB on disk after extraction. Everything downloads on first run; no profiles, no opt-ins.
npm install -g multilingual-dictionary-mcp{
"mcpServers": {
"multilingual-dictionary": {
"command": "npx",
"args": ["-y", "multilingual-dictionary-mcp"]
}
}
}On first run, the server downloads its data bundle (~5.4 GB compressed → ~15.6 GB on disk) from multilingual-dictionary-mcp-data.nyc3.cdn.digitaloceanspaces.com into ~/.cache/multilingual-dictionary-mcp/ (overridable via MDM_DATA_DIR). Each artifact is gzip-decompressed during the stream and SHA-256 verified. Subsequent runs are instant.
dictionary_synonyms, dictionary_antonyms, dictionary_hypernyms, dictionary_hyponyms, dictionary_meronyms, dictionary_holonyms, dictionary_derived_from, dictionary_etymologically_related, dictionary_used_for, dictionary_capable_of, dictionary_at_location, dictionary_translate, dictionary_all_relations
Tool | What it does |
| Semantic neighbors via cosine similarity — much denser than ConceptNet RelatedTo |
| Explicit embedding lookup with optional cross-lingual filter |
| Multilingual "means approximately" via Numberbatch (works in any of the 78 covered languages) |
Tool | What it does |
| Definitions + IPA + etymology grouped by language |
| Concatenated short definition |
| Etymology text |
| IPA pronunciations |
| Prefix-search the corpus |
| Random word entry |
Tool | What it does |
| Perfect or near rhymes |
| Homophones / soundalikes |
| Spelling pattern ( |
| Autocomplete prefix |
> What are the synonyms of "feliz" in Spanish?
[uses dictionary_synonyms with language: "es"]
> Translate "happiness" into Hebrew
[uses dictionary_translate with language: "en", targetLanguage: "he"]
> What's the etymology of "serendipity"?
[uses dictionary_etymology]
> Find words semantically similar to "café" across languages
[uses dictionary_semantic_neighbors]
> What rhymes with "orange"?
[uses dictionary_rhymes with perfect: false]npm install
npm run build
npm test # 42 tests, all run without local dataSource layout:
src/
├── index.ts # MCP server entry, blocks on first-run install
├── data/
│ ├── paths.ts # CDN base, profile selection, data dir
│ ├── installer.ts # First-run downloader + SHA-256 verification
│ └── local-store.ts # SQLite + Numberbatch matrix wrappers
└── tools/
├── relations.ts # 13 ConceptNet relation tools + 1 embedding tool
├── definitions.ts # 6 Wiktextract definition tools
└── english.ts # 5 CMU dict + Numberbatch toolsTotal: 24 tools.
The CDN ships official builds, but the entire pipeline is in data-pipeline/ — see its README. You can rebuild from upstream sources, ship to your own bucket, and override MDM_CDN_BASE.
CI runs build + tests on Node 20 & 22 for every push and PR.
Publish runs on GitHub Release publish, pushes to npm and GitHub Packages.
MIT © Eyalm321
All edges (every relation type) for a word in one call from offline ConceptNet.
Inputs: word (required), language, limit.
Get antonyms (opposites) for a word in any language using offline ConceptNet.
Inputs: word (required), language, limit.
Typical locations (AtLocation) via offline ConceptNet. E.g. book -> library.
Inputs: word (required), language, limit.
Typical actions (CapableOf) via offline ConceptNet. E.g. dog -> bark.
Inputs: word (required), language, limit.
Derivation relations via offline ConceptNet DerivedFrom.
Inputs: word (required), language, limit.
Return etymology text for a word from the offline Wiktextract corpus.
Inputs: word (required), language.
Holonyms (wholes that contain this word) via offline ConceptNet reverse PartOf.
Inputs: word (required), language, limit.
Hypernyms (broader concepts) via offline ConceptNet IsA. E.g. dog -> mammal.
Inputs: word (required), language, limit.
Hyponyms (narrower concepts) via offline ConceptNet IsA. E.g. dog -> poodle.
Inputs: word (required), language, limit.
Manually kick off (or re-trigger) the offline data download. Returns immediately while the install runs in the background. Idempotent — calling repeatedly while a download is in progress is a no-op. Use dictionary_status to track progress.
Inputs: no named parameters.
Look up definitions for a word from the offline Wiktextract corpus (4,755 languages, 10.5M entries). Returns part-of-speech, definitions per sense, IPA, etymology, and translations grouped by language.
Inputs: word (required), language, limit.
Find words/phrases meaning approximately the same as the input via offline Numberbatch embedding cosine. Multilingual — works in any of the 78 languages Numberbatch covers.
Inputs: query (required), language, limit.
Meronyms (parts/components) via offline ConceptNet PartOf. E.g. car -> wheel.
Inputs: word (required), language, limit.
Return IPA pronunciation(s) for a word from the offline Wiktextract corpus.
Inputs: word (required), language.
Return a random word entry from the offline Wiktextract corpus, optionally filtered to a language.
Inputs: language.
Find words that rhyme with the input word. English only — backed by the offline CMU Pronouncing Dictionary.
Inputs: word (required), limit, perfect.
Prefix-search the offline Wiktextract corpus for words matching a query. Useful when you don't know the exact spelling.
Inputs: query (required), language, limit.
Embedding-based nearest neighbors via Numberbatch. Returns words from the same OR a different language, sorted by cosine similarity. Multilingual (78 languages).
Inputs: word (required), language, targetLanguage, limit.
Find English words that sound similar to the input (homophones / soundalikes) via offline CMU dict.
Inputs: word (required), limit.
Find English words matching a spelling pattern via offline CMU dict. '?' = any single letter, '*' = any zero+ letters. Useful for crosswords.
Inputs: pattern (required), limit.
Get the current state of the offline data installation. Always available — even before the bundle has finished downloading. Returns: state (pending/downloading/ready/failed), per-artifact progress, total bytes downloaded, data directory, CDN URL, manifest version, and any error.
Inputs: no named parameters.
Get autocomplete suggestions for a partial English word via offline CMU dict (shortest matches first).
Inputs: prefix (required), limit.
Get a brief plain-text summary of a word's senses from the offline Wiktextract data. Concatenates the first definition of each sense.
Inputs: word (required), language.
Get synonyms for a word in any language using the offline ConceptNet 5.7 SQLite (24M edges, 80+ languages).
Inputs: word (required), language, limit.
Translate a word to another language via offline ConceptNet's cross-lingual Synonym links. Returns the same concept expressed in different languages.
Inputs: word (required), language, targetLanguage, limit.
Typical uses (UsedFor) via offline ConceptNet. E.g. knife -> cutting.
Inputs: word (required), language, limit.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 28 tools
Each tool targets a specific linguistic relation or operation with clear, distinct descriptions. Even potentially overlapping tools like synonyms, means_like, and related are differentiated by their underlying data sources (ConceptNet vs Numberbatch embeddings).
All tools follow a consistent 'dictionary_' prefix followed by a descriptive snake_case name. The pattern is uniform with no mixing of conventions or styles.
28 tools is on the higher end but justified by the comprehensive coverage of dictionary features (definitions, etymology, pronunciation, semantic relations, search, data management). It remains manageable and focused.
The tool surface covers a wide range of dictionary and lexical relations: definitions, pronunciations, etymology, synonyms, antonyms, hypernyms, hyponyms, meronyms, holonyms, translations, rhymes, spelling, autocomplete, random access, and data installation status. No obvious gaps for the stated purpose.
MCP server for Chinese dictionary lookup: character definitions, pinyin, stroke order and word
Wiktionary MCP — multilingual dictionary via Wikimedia REST + MediaWiki Action API
Dictionary MCP — wraps Free Dictionary API (free, no auth)
MCP server for Chinese idioms (chengyu): lookup by word, pinyin or meaning with explanation and