Privacy and Protected Content

UnimeType sends text only when the user invokes Convert, Polish, or Explain and has granted AI-processing consent. Ordinary typing does not automatically send requests.

Locally masked objects

Before a request, UnimeType replaces these recognized values with local tokens:

• Backtick code spans and fenced code blocks
• HTTP, HTTPS, and www. URLs
• Email addresses
• Recognized Unix and Windows paths
• Markdown link destinations
• Values beginning with recognized sk- or AIza key patterns
• PEM, OpenSSH, and PGP private-key blocks
• The remainder of a line following a recognized secret label

Recognized labels include password, passwd, pwd, passcode, PIN, API key, x-api-key, provider API key, access token, auth token, refresh token, client secret, authorization, private key, seed phrase, recovery phrase, mnemonic, wallet seed, secret, token, 密码, 口令, 私钥, 助记词, and 恢复短语.

For a labelled secret, the complete remainder of that line is masked. This avoids exposing a password merely because it contains spaces or punctuation.

Restoration contract

The configured model sees placeholder tokens rather than the protected values. Every target placeholder must appear exactly once in the response. A missing, duplicated, reordered, or structurally moved token causes validation or local restoration to fail; UnimeType keeps the original host text.

The privacy-lock indicator appears for recognized secrets. Code, URLs, email, paths, and link targets are also protected, but do not by themselves trigger the secret indicator.

Context sent by Convert

For a newly appended suffix in an accepted paragraph, Convert may include up to 100 preceding characters as read-only context. Protected values in both the context and writable target are masked independently. The model is instructed to return only the writable target.

Local masking is pattern-based and cannot identify every possible secret. Do not submit a value when you do not want any part of the surrounding text sent to the configured service.