Purpose: anonymize-mcp is an independent community project that wraps the public NLP tools of LINDAT/ÚFAL MFF UK (NameTag, MasKIT, PONK, UDPipe, Korektor, Charles Translator) as an MCP server, mainly used for personal information anonymization in GDPR scenarios, as well as entity recognition, readability, and proofreading processing of Czech legal texts.
Main capabilities: provides 6 MCP tools—extract_entities (multilingual NER with automatic detection of 35 languages), anonymize (production-grade pseudonymization, with 80+ PII patterns, covering 9 industries and international rules for the United States, United Kingdom, Germany, France, Italy, Spain, Poland, Russia, India, etc., EU 28 country tax numbers, 30+ country IBANs, cryptocurrency and API tokens), analyze_morphology (UDPipe 961 models), check_readability, correct_text (spelling and diacritics), translate_text (8 languages, 17 direct translation pairs and English pivot).
Usage: install via pip install anonymize-mcp (or install from source), register it as a standard stdio server in an MCP client, and then it can be called. The materials list configuration methods for Claude Code, older Claude Desktop, OpenAI Codex CLI, Cursor, etc.; the new Claude Desktop (Microsoft Store/appx) only supports remote HTTP MCP and cannot add such local stdio servers. On Windows, if the anonymization executable is not in PATH, the configuration must use the full path.
Zero-egress local mode: after setting ANONYMIZE_MCP_LOCAL=1, anonymize and extract_entities run completely offline (local regex plus CNEC 2.0 NameTag), and text does not leave the local machine, making it suitable for sensitive data such as legal and medical data; the other tools are rejected by default with a prompt, and can be explicitly allowed via ANONYMIZE_MCP_LOCAL_ALLOW_CLOUD=1. Models can be pre-downloaded via python -m anonymize_mcp.local_backend (SHA-256 verification), and combined with ANONYMIZE_MCP_NO_DOWNLOAD=1 to achieve a no-network-connection environment.
Limitations: cloud mode sends text to external servers, and sensitive data should first be processed through anonymize. Local NameTag 1 is inferior to cloud NameTag 3, and lacks the MasKIT rule engine. The project is for non-commercial use only; the NameTag and UDPipe models use CC BY-NC-SA; the LINDAT API is free for academic and personal use, while bulk, paid, or production traffic requires explicit permission from the authors and the API operator. This project is an unofficial community project by an independent developer and is not operated or endorsed by ÚFAL MFF UK; the MIT license covers only this project, and the four underlying tools each have their own independent licenses.
