DeepL Translate (confidence-gated fallback)

A thin wrapper around the DeepL API. The point of this skill is the trigger logic, not the API call: translate with DeepL when your own translation might be wrong and being wrong matters.

When to reach for DeepL instead of translating yourself

Translate it yourself when the text is everyday prose and you are confident. Call DeepL when any of these is true:

• Proper nouns / brand / product names that have official localized forms
• Domain terminology — legal, medical, financial, technical specs, patents
• Ambiguous source where one word maps to several target words and context doesn't disambiguate
• Idioms / fixed expressions that don't translate literally
• Low-resource or distant language pairs where your training signal is thin
• High-stakes output — anything the user will publish, sign, or send to a third party, where a subtle error is costly
• The user explicitly asks for DeepL.

If you're confident and the cost of a minor error is low, just translate directly — don't burn an API call.

Prerequisite: API key

The key is read from the DEEPL_API_KEY environment variable (never hardcode it). Default endpoint is the Free tier host api-free.deepl.com; for Pro, set DEEPL_API_HOST=api.deepl.com.

Set it once:

# macOS / Linux — add to ~/.bashrc or ~/.zshrc to persist
export DEEPL_API_KEY="your-deepl-auth-key-here"

# Windows (persistent, user scope) — then open a NEW shell
setx DEEPL_API_KEY "your-deepl-auth-key-here"

Verify in a fresh shell: echo $DEEPL_API_KEY (bash) / $env:DEEPL_API_KEY (PowerShell).

If DEEPL_API_KEY is unset, stop and tell the user to set it — do not invent a key.

How to call it

Use the bundled Node helper (cross-platform, Node 18+). --source is optional; omit it to let DeepL auto-detect. --text also answers to -t, and --target/--source to --target-lang/--source-lang. Every flag takes its value either as the next argument or after an = — the = form is the one that can carry text beginning with --.

node "{SKILL_DIR}/translate.mjs" --target ZH \
    --text "The plaintiff filed a motion to compel discovery."

With an explicit source language:

node "{SKILL_DIR}/translate.mjs" --source JA --target EN-US --text "持分会社"

The script prints only the translated text on stdout, so $(...) around it captures the translation and nothing else. Failures go to stderr as one ERROR: line plus a non-zero exit code. Multiple lines / paragraphs are preserved. The request is capped at 60 s, so a stalled endpoint (throttled, blackholed, captive portal, dead proxy) prints an ERROR: naming the timeout instead of hanging — never impose your own deadline on top.

Language codes

These cover the usual work without a round trip. They are a starting list, not a gate: a language missing here is still probably supported, so query the endpoint below rather than deciding it cannot be done and translating it yourself.

• Core (source + target): AR BG CS DA DE EL EN ES ET FI FR HU ID IT JA KO LT LV NB NL PL PT RO RU SK SL SV TR UK ZH
• Extended (among ~100, incl. many low-resource ones): AF BN FA HE HI HR HY KA ML MR MS MY NE PA SW TA TE TH UR VI YUE ZU … and many more.
• Target-only regional variants — prefer these when precision matters:
  • Chinese: ZH-HANS (Simplified), ZH-HANT (Traditional) — better than bare ZH
  • English: EN-US, EN-GB
  • Portuguese: PT-BR, PT-PT
  • Spanish: ES-419 (Latin American)
  • FR-CA (Canadian French), DE-CH (Swiss German)

For the authoritative, always-current list, query the v3 languages endpoint (/v2/languages is superseded). The resource query parameter is required — omitting it returns 400: GET https://api-free.deepl.com/v3/languages?resource=translate_text with the Authorization: DeepL-Auth-Key <key> header (other resource values: translate_document, translation_memory, voice, write, glossary, style_rules). Each entry carries usable_as_source and usable_as_target separately, so one call tells you both directions. Or see https://developers.deepl.com/docs/getting-started/supported-languages

Do not treat any listing as the gate. /v2/languages?type=target does not list EN, PT or YUE as targets, yet /v2/translate accepts all three and translates into them; v3 lists more languages than v2 and reports en/pt as usable targets, agreeing with the behaviour. So a code missing from a listing is not evidence of no support — if it matters, make the call and read the error. A code DeepL really rejects comes back as HTTP 400 — Bad request. Reason: Value for 'target_lang' not supported. It names the field rather than echoing the code, so it tells you the language was the problem and clears the text.

Endpoint note: /v3/languages returns BCP 47 codes (en-US, pt-BR, zh-Hant) while this skill writes them uppercase, but you can feed either straight into a translate call — /v2/translate matches target_lang case-insensitively and honours the regional variant both ways (zh-Hant and ZH-HANT both return Traditional). Text translation stays on /v2/translate: current, not deprecated, and with no v3 equivalent.

Workflow inside a task

1. You're translating something and hit a passage you're unsure about.
2. Confirm DEEPL_API_KEY is set (the script checks and errors clearly if not).
3. Run translate.mjs for the uncertain passage (or the whole text).
4. Use DeepL's output. If it conflicts with your own read, surface both to the user and explain the discrepancy rather than silently picking one.

Glossary / terminology consistency (optional)

For repeated terminology, pass --glossary as a term=translation;term=translation list. It is a literal find-and-replace run over DeepL's output, case-sensitively and substring-wide — API=接口 also rewrites APIs — so keep terms unambiguous and long enough not to collide, and a term cannot contain ;. Whitespace around a term is trimmed. A pair with no = is skipped with a WARNING: line on stderr rather than quietly doing nothing — the translation still arrives on stdout, so a caller capturing it is unaffected. For real glossary support DeepL has a Glossary API; ask the user if they need it and we can extend the script.

Notes

• The script does not track quota; a 456 response means it is exhausted. What that quota is depends on the account's plan, so read it from the DeepL console rather than from here.
• 403 means a bad/missing key.
• A timeout means the host accepted the connection but never replied — a network/reachability problem, not a bad key or bad input. Retrying immediately usually reproduces it.
• For DeepL Pro, set DEEPL_API_HOST=api.deepl.com (default is api-free.deepl.com).