Datamuse Mcp Server

Overview

Word finding for writing, editing, puzzles, and naming, powered by the Datamuse API. Find a rhyme with a particular meaning, explore lexical alternatives, or combine meaning, sound, spelling, and relation constraints.

Tools

Capability reference

datamuse_find_rhymes _tool

• Required word; rhyme_type is exact (default) or near, never blended; optional means_like and spelled_like narrow the same search. English only
• Shared search controls: limit (default 20), offset (default 0), include_definitions, include_pronunciation, and include_ipa (metadata options default false). Rows carry whatever metadata Datamuse supplies

datamuse_find_synonyms _tool

• Required word; relation is synonyms (default), antonyms, broader (direct hypernyms), or narrower (direct hyponyms). English only
• Same shared search controls as rhymes; statistical associations are the associated relation of datamuse_find_words

datamuse_find_words _tool

• At least one of means_like, sounds_like, spelled_like (upstream spelling patterns), or up to 14 relations: [{ relation, word }], all conjoined; repeated relations are kept. left_context, right_context, and up to five topics, each a single word, affect ordering only
• Relations: exact_rhyme, near_rhyme, synonyms, antonyms, broader, narrower, parts, wholes, nouns_modified_by, adjectives_describing, followers, predecessors, associated, homophones, consonant_match; datamuse_list_reference gives their directions
• vocabulary is en (default) or es; Spanish takes meaning, sound, and spelling constraints, not relations. Rejections are typed: missing_constraint, unsupported_vocabulary_relation, invalid_context_hint

Spelling patterns are interpreted by Datamuse using supported OneLook syntax:

For anagrams, call datamuse_find_words with {"spelled_like":"//listen"}. +letters restricts allowed letters and -letters excludes them; neither requires every listed letter.

datamuse_get_word_info _tool

Supply 1–100 nonblank words, each at most 200 characters. The first ten are processed sequentially through the shared request budget; later entries return processing_cap. One 20-second deadline includes all requests, retries and pacing. Completed results survive item failures; pending entries return deadline_exceeded. Caller cancellation stops the call.

The same include_definitions, include_pronunciation, and include_ipa options apply. Each indexed outcome contains trimmed caller input and either a complete raw row, a failure reason with recovery, or an unprocessed reason. Only a query-tagged echo matching the supplied spelling under Unicode lowercase is accepted: Ice Cream can return raw row.word='ice cream', while corrections are never substituted. Missing metadata stays absent, and duplicates keep their indices. Lookup output is complete on both MCP surfaces and has no search byte slicing.

datamuse_list_reference _tool

• topic is all (default), relations, metadata, combinations, or limitations
• Returns static sections: [{ topic, text }] without a network request

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Datamuse-specific:

• Live queries to Datamuse's fixed HTTPS /words endpoint, with no local mirror or stored queries
• One request budget shared by every tool and tenant: one start per second, one request in flight, a bounded queue, and one retry. It is process-local, so it does not account for other processes or other users of the same egress address
• Results are corpus-derived suggestions with sense-dependent, sometimes guessed metadata (parts of speech, syllables, f: frequency per million words). They do not establish dictionary completeness or poetic suitability

Agent-friendly output:

• Effective-query echo and honest caps: every search returns effectiveQuery, returnedCount, limitReached, and truncated, plus relevant notices and nextOffset. Positive safe-integer limits above 1000 are clamped with requestedLimit and effective limit echoed. Invalid integers reject before I/O
• Whole-row navigation: offset accepts 0–999. Searches fetch max=min(1000, offset+limit) and select complete rows under 64,000 bytes per serialized MCP surface, including escaping and enrichment. A local byte cut sets truncated=true; re-run with offset=nextOffset when supplied. One oversized row is returned intact with a soft-budget notice and offset progress. API-cap notices are separate from byte cuts. Rankings can change between calls, so offsets are best-effort navigation with possible duplicates or gaps, never snapshots, stable pagination, totals, or proof of exhaustion. No continuation beyond index 999
• Phrase relation recovery: phrase targets stay accepted and unchanged. When a relation query returns empty, guidance explains limited phrase coverage and suggests means_like or sounds_like; rhyme guidance also suggests explicitly trying the final word
• Optional pronunciation: include_pronunciation appends r to md; include_ipa implies pronunciation and adds ipa=1. Raw pron: and ipa_pron: tags may both occur. ARPAbet stress digits mean 0 unstressed, 1 primary, 2 secondary; pronunciations may be guessed, sense-dependent, or absent
• Lossless rows: word plus whatever score, numSyllables, tags, defs, defHeadword, and other fields Datamuse sent, unchanged and in upstream order, identical in structuredContent and the fenced JSON text. score ranks results within one query only
• Typed failures: declared reasons for access errors, rate limits (upstream_rate_limited and pacer_shed, with retryAfter), malformed upstream data, and invalid input, each with a recovery hint

Access, privacy, and attribution

Datamuse currently documents keyless access up to 100,000 requests per day and asks customer-facing applications to contact it. It has announced manually issued API keys for early 2027, with conflicting dates: January 1 in its usage section and February 1 in its banner. No key format is documented yet, so this server uses the keyless API, and access may be interrupted at the switch.

Queries are sent to Datamuse, which keeps a temporary log of them until the end of the following day. Server payload logging and telemetry are optional; review their settings before sending sensitive text.

Word results and lexical metadata come from Datamuse; see its documentation for data sources and attribution requirements. The Apache-2.0 license covers this server's code, not the upstream data.

Getting started

Add the following to your MCP client configuration file.

{  "mcpServers": {    "datamuse-mcp-server": {      "type": "stdio",      "command": "bunx",      "args": ["@cyanheads/datamuse-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info"      }    }  }}

Or with npx (no Bun required):

{  "mcpServers": {    "datamuse-mcp-server": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@cyanheads/datamuse-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info"      }    }  }}

Or with Docker:

{  "mcpServers": {    "datamuse-mcp-server": {      "type": "stdio",      "command": "docker",      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/datamuse-mcp-server:latest"]    }  }}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No Datamuse credentials are currently required; see Access, privacy, and attribution for the announced key transition.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/datamuse-mcp-server.git

2. Navigate into the directory:

cd datamuse-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env# edit .env for transport, authentication, logging, or telemetry overrides

Configuration

See .env.example for every server setting and the common framework overrides.

Running the server

Local development

• Build and run the production version:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:http# orbun run start:stdio

• Run checks and tests:

  sh

  bun run devcheck  # Lints, formats, type-checks, and morebun run test      # Runs the test suite

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for logging, ctx.state for storage
• Register new tools and resources in the createApp() arrays in src/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.