The dictionary holds 1,883 entries and 1,263 of them carry a citable source — Wilhite on GIF, Crockford on JSON, RFC 7519 on JWT — with each entry tagged creator-clarified, community-consensus or contested, so a genuinely disputed reading comes back as disputed instead of as one arbitrary answer. Words outside the dictionary return an error, which is why the answers you do get are worth something.
A community-maintained pronunciation dictionary for developer project, product and jargon names, served over MCP. Ask for a word and you get how it is said, how the reading was settled, and audio you can play.
- The canonical reading of a developer word: IPA, a respelling of how engineers say it, a playable MP3 audio_url, alternate readings, the source citation and editorial notes
- Substring search over the dictionary's word and category fields, returning word, slug, category and confidence
- Entries listed by category — product, project, cli-tool, tool, cs-term, acronym, abbreviation, person
- A confidence tag on every entry: creator-clarified, community-consensus or contested
- An error for a word that is not in the dictionary, rather than an invented pronunciation
The server lives in the repository's mcp-server subdirectory and ships to PyPI as pronounce-mcp; the documented launch runs it with uvx straight from that subdirectory, and there is an OCI image at ghcr.io/anzy-renlab-ai/pronounce-mcp. Speaks stdio. It reads the public dictionary at pronounce.renlab.ai, so there is no account and no key. The CLI and editor extensions in the same repository are separate from the server.
One command — uvx pronounce-mcp