create_api_key
ChatGPTUse this when the user wants to call a Hopper model from their own app, script or agent, or asks for a Hopper API key. Creates a key on the signed-in account and returns its name, its first characters and a one-time link the user opens to copy it (the key itself is never in this result), plus the environment line and client snippet for the model. Ask the user to store it as HOPPER_API_KEY in server-side secrets, such as a gitignored .env; don't ask them to paste it into the chat. Keys are managed at https://withhopper.com/console/keys.
generate
ChatGPTUse this when the user wants to try, test or compare open-source LLMs on Hopper on their own prompt or examples: DeepSeek V4.1 Flash Oct 16 release (deepseek-v4.1-flash-1017, the newest DeepSeek: use it for "the new DeepSeek"), DeepSeek V4.1 Flash Sep 10 release (deepseek-v4.1-flash) and Gemma 4 31B (gemma-4-31b). For example "how is the new DeepSeek model for my use case", "try this prompt on DeepSeek", or "compare Gemma 4 and DeepSeek on this". Runs the prompt on each model in models side by side and returns each answer with time to first token, tokens per second, token counts and cost. Pass the user's real prompt, system prompt or examples, not a summary. Works without an account (up to 2048 output tokens per call); when the user is signed in, usage is billed to their Hopper account. Not for writing your own reply to the user.
get_account
ChatGPTUse this when the user asks about their Hopper credits, spend, usage or API keys. Shows the signed-in account: credits remaining, total spend, request and token counts, and API keys (names and first characters only; secrets are never readable). Read-only: it can't delete, revoke or reveal keys.
get_integration_guide
ChatGPTUse this when the user wants to call a Hopper model from their own code: an app, script, backend or agent. Returns tested client code for Hopper's OpenAI-compatible API: openai-sdk for any app or agent (Python or JavaScript), or livekit, pipecat or vapi for those frameworks. Each guide uses an HTTP/2 client and a prompt-cache warm-up.
list_models
ChatGPTUse this when the user asks which models Hopper serves, whether a model is on Hopper, what they cost, or which model to use. Lists the LLMs (DeepSeek V4.1 Flash including its Oct 16 release, Gemma 4 31B), text-to-speech (Qwen 3 TTS) and speech-to-text (Nemotron ASR) models with context length, prices (what the API bills) and whether each is available now.
list_voices
ChatGPTUse this when choosing a voice for speak, or when the user asks what voices Hopper has. Lists the featured voices ready for text to speech, with id, name, language, gender and a short description. Works without an account.
speak
ChatGPTUse this when the user wants text read aloud, a voiceover or narration (for a launch video, demo, ad, podcast intro or announcement) or an audio file of speech (Qwen 3 TTS). Makes the speech track only, not video, music or sound effects. Shows an audio player and returns 24 kHz mono WAV audio with a playable HTTPS link (valid 10 minutes, 3 plays; download it to keep the file). Up to 4000 characters; longer text is split at sentence ends. Works without an account, limited to about 1 minute of audio per call and 10 minutes a day; when the user is signed in, usage is billed to their Hopper account without those limits. Do not use for music or sound effects.
transcribe
ChatGPTUse this when the user wants speech in a recording turned into text, or the agent needs to hear it: transcripts, captions or subtitles, checking what is said in a voiceover or video soundtrack, meeting notes, call summaries or interview quotes. English speech to text with word timestamps (Nemotron ASR). Takes WAV, MP3, M4A, Ogg/Opus, FLAC or WebM (or raw PCM16 mono with sample_rate): a file the user attached in the chat (pass it as audio_file as it is: don't download, convert or transcribe it yourself first), a public https URL (audio_url, up to 25 MB), or base64 (audio_base64, up to 3 MB); up to 5 minutes of audio per call. Agents with a local file and a shell (not chat attachments) can call it with no audio to get a one-time upload URL for curl (up to 4 MB). Works without an account, limited to 30 minutes of audio a day; when the user is signed in, usage is billed to their Hopper account.
