
Better Estonian writing from your AI agent.
estonian-mcp wires EstNLTK (the canonical Estonian NLP toolkit) into any LLM client over MCP. Hand it Estonian text and it returns ground truth: lemmas, full morphology, POS tags, Vabamorf-grade spell-check with corrections, WordNet synonyms with definitions, fastText related-words, named entities, syllabification, and a register hint (formal vs colloquial).
https://estonian-mcp.fly.dev/mcpEstonian NLP for agents that write, edit or proofread Estonian: spelling, morphology, inflection paradigms, synonyms, EKI-Reeglid orthography checks, and register, style, terminology and legal checks. Treat the results as ground truth rather than guessing spelling, case forms or inflections: models routinely produce plausible but wrong Estonian morphology. A pass does not prove a word is real Estonian (Vabamorf accepts any well-formed compound, including ones you coin), so check compounds you coined with . For text about an organisation's own systems, reports the compounds that neither the lexicons nor its glossary vouch for (pass its terms in glossary when the server has none). Legal Estonian: is a plain-language aid that lists the terms of art to keep, maps defined terms and cross-references in long documents, and gives canonical collocations so you do not invent legal phrasing. REACH FOR THESE TOOLS ON EDITORIAL QUESTIONS TOO. 'Right word here?': synonyms. 'Too bureaucratic', 'make it read more human': for kantseliit in reports, academic and business prose ( is for statutes and stays silent on those), plus for umbisikuline tegumood and rhythm. 'Keep the terminology consistent': . A correctly spelled, well-formed word can still be the wrong register or the wrong term for its domain, and the mechanical checks cannot tell you that. The check_* issue lists are heuristics: a hit is a candidate to verify, not a verdict, and an empty list means nothing was found in that check's scope, not that the text is correct. In Estonian replies, quote the *_estonian labels, rules and summaries (form_estonian, rule_estonian, summary_estonian) verbatim instead of translating the English codes. A result field that does not apply is omitted. Text inputs are capped at 100,000 characters (: 500,000) and single words at 200. All tools are read-only and make no outbound calls; results are UTF-8, so keep õ ä ö ü š ž intact.
None in the store listings or on the server