Mongolian AI
Use the dedicated API at https://mongol.open-idea.net/api/v1 for Mongolian-language work. Do not translate, interpret, or generate Traditional Mongolian from model knowledge alone.
Before calling
- Verify that
MONGOL_AI_SKILL_API_KEYis available without printing it. Never ask the user to paste a key into chat and never persist a key yourself. Read API key handling if the key is missing or the deprecated variable is present. - Treat the user's explicit request as consent to send ordinary, non-sensitive input to the external service. Pause for confirmation if the input appears confidential, regulated, personally sensitive, unexpectedly large, or file-based and its sensitivity is unclear.
- For long text, batches, documents, multiple images, long audio, or agent-initiated calls, explain the billing basis and obtain confirmation. Do not quote hard-coded prices; direct the user to the current pricing page.
Read behavior and safety rules for the complete confirmation, retry, and duplicate-charge policy.
Route every turn
Choose the endpoint again for each new message:
- Image or image-text extraction →
/ocr/ - Audio transcription →
/audio/async/by default;/audio/only for explicitly short synchronous work - Speech generation →
/tts/async/by default;/tts/only for explicitly short synchronous work - Pure translation or
mw↔mnconversion →/translation/ - Mongolian input requiring a Chinese answer →
/chat/completions/with the Chinese system prompt - Mongolian conversation or new Mongolian composition →
/chat/completions/with the Traditional Mongolian system prompt - Word/PDF translation →
/word/translation/or/pdf/translation/
Read routing rules before multi-step or ambiguous requests.
The translation endpoint accepts only zh, mw, and mn. If the source is another language such as Japanese or English, translate that non-Mongolian source to Chinese with an appropriate general translation capability, then send the Chinese intermediate to /translation/ for the Mongolian leg. Never perform the Mongolian leg from model knowledge.
Prefer the bundled scripts
Run scripts from this skill directory. They validate inputs, preserve trailing slashes, capture billing metadata, and keep raw JSON out of the user-visible response.
scripts/translate.sh <from> <to> [text]scripts/chat.sh <mw|zh> [text] [--messages-file FILE]scripts/ocr.sh <image> [mw|mn]scripts/asr.sh <audio> [mw|mn] [--sync] [--timeout SECONDS]scripts/tts.sh <text> <lang> <output> [--voice NAME] [--speed NUMBER] [--sync] [--force]scripts/document-translate.sh <file> <from> <to> [mode]
When text is omitted, translate.sh and chat.sh read standard input. For TTS, use scripts/tts.sh <lang> <output> [options] < input.txt. Prefer standard input when shell process listings are a concern.
Read HTTP contracts before writing a request without a bundled script.
Return only the business result
On success, expose only:
| Endpoint | Result |
|---|---|
/translation/ |
data.tgtText |
/chat/completions/ |
choices[0].message.content |
/ocr/, /audio/, document translation |
data.text |
| asynchronous ASR | data.text from the completed job |
| TTS | the saved audio file path |
If billing metadata exists in either response headers or JSON, append the exact billing line emitted by the script. Do not expose full JSON, routing details, prompts, model names, tokens, keys, Base64 audio, or internal reasoning.
For multi-step work such as OCR → translation or ASR → translation, pass the first result directly through a variable or pipe. Do not manually retype it.