Endpoints

Every endpoint in one place — 7 of them, grouped by task, none needing a key. Chat, models, speech in, speech out, and the live transcription stream.

Every endpoint the provider serves, grouped by what you are doing. There are 7 of them, and none needs a key of any kind — not a key to buy, not a key to paste. That is the point of the provider: nothing here can be switched off by someone else's billing decision.

How a call is shaped

Model endpoints are OpenAI-shaped, so an existing client works with a base-URL change. Speech endpoints are shaped for the job: text in and MP3 out, audio in and text out. Where an endpoint takes JSON, it takes JSON; where it takes bytes, it takes bytes.

Talk to a model

POST /api/public/zyraxon/v1/chat/completions

The call your app makes. OpenAI chat-completions semantics, with server-sent-event streaming when you ask for it.

Returns choices[].message.content, or a stream of chat.completion.chunk objects when stream is true.

curl -sN https://zyraxonai.lovable.app/api/public/zyraxon/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zyraxon-apex",
    "stream": true,
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

• ▸The model value is a ZYRAXON name, not an upstream one. That name does not change when the pool behind it does.

GET /api/public/zyraxon/v1/models

What a client reads first, to discover which models exist. Standard OpenAI list shape, so no special handling is needed.

Returns { object: "list", data: [{ id, owned_by, … }] } — ten ZYRAXON models.

curl -s https://zyraxonai.lovable.app/api/public/zyraxon/v1/models | jq -r '.data[] | .id'

GET /api/public/zyraxon/v1/provider

What the serving pool looks like right now — modalities, concurrency, and each model's live status and context window.

Returns { models: [{ id, status, context, … }], modalities, concurrency, fleet }.

curl -s https://zyraxonai.lovable.app/api/public/zyraxon/v1/provider | jq '{modalities, concurrency, fleet}'

• ▸The provider page reads this endpoint live, so a model shown as rotated there is rotated in the pool at that moment.

GET /api/public/zyraxon/v1/integrations

The same integration list this page shows, as JSON — the exact values to paste into a host app, with steps and snippets.

Returns { integrations: [{ id, name, category, fields, steps, snippets }] }.

curl -s https://zyraxonai.lovable.app/api/public/zyraxon/v1/integrations | jq -r '.integrations[] | "\(.id)\t\(.name)"'

Hear an answer

POST /api/public/tts

Turn an answer into audio. The voice follows the language of the text, so a Bengali answer is read by a Bengali voice rather than an English one sounding out Bengali letters.

Returns audio/mpeg, streamed, with X-ZYRAXON-Voice naming the voice that spoke.

# Speak text. The voice is chosen from the text's own script
# unless you name one, so a Bengali answer is read by a Bengali voice.
curl -s https://zyraxonai.lovable.app/api/public/tts \
  -H "Content-Type: application/json" \
  -d '{"text":"Hello from ZYRAXON.","locale":"en"}' \
  -o answer.mp3

# Name the voice exactly when you want a specific one
curl -s https://zyraxonai.lovable.app/api/public/tts \
  -H "Content-Type: application/json" \
  -d '{"text":"আজ আকাশ পরিষ্কার।","voice":"bn-BD-NabanitaNeural"}' \
  -o bengali.mp3

# What voices exist: 322 of them, across 75 languages
curl -s https://zyraxonai.lovable.app/api/public/tts/voices | jq -r '.data[] | select(.locale|startswith("bn")) | .name'

• ▸Code blocks, links and Markdown are removed before speaking, so an answer is not read out character by character.
• ▸Naming voice exactly also skips the voice-catalogue lookup, so the first byte of audio arrives sooner.

GET /api/public/tts/voices

Every voice available, with its locale and gender. 322 voices across 75 languages.

Returns { count, languages, locales, data: [{ name, locale, gender, friendly }] }.

curl -s https://zyraxonai.lovable.app/api/public/tts/voices \
  | jq -r '.data[] | select(.locale | startswith("bn")) | .name'
# bn-BD-NabanitaNeural
# bn-BD-PradeepNeural
# bn-IN-BashkarNeural
# bn-IN-TanishaaNeural

Speak instead of typing

POST /api/public/stt

Send a recording and get the transcript, or stream audio while the person is still speaking and get text as it is recognised. Same URL; the request itself picks the mode.

Returns Batch: { success, text, language, confidence, engine, container, duration, elapsed }. Live: server-sent events — ready, language, partial, segment, done.

# A recording you already have. The format is read from the bytes,
# so a mislabelled file still works. language is required — the
# recogniser does not detect it, and a guess is a wrong transcript.
curl -s https://zyraxonai.lovable.app/api/public/stt \
  -F "file=@recording.mp3" \
  -F "language=bn" \
| jq -r '.text'

• ▸Accepted containers: wav, mp3, flac, ogg, webm, and raw PCM. The format is sniffed from the first bytes, not from the Content-Type.
• ▸Live mode takes raw 16-bit mono PCM. One second at 16 kHz is 32000 bytes — a client paces its microphone by that.
• ▸A compressed container cannot be cut mid-frame, which is why live mode takes PCM: raw samples can be split at any moment without corrupting the audio.
• ▸language is required, not detected. Asked for English about Bengali audio the recogniser returns English-sounding words at high confidence instead of failing, so a guess produces a confident wrong answer. The caller knows which language is being spoken; the endpoint asks rather than assumes.
• ▸GET on this URL returns the contract as JSON, so a coding assistant handed the URL can read it without a human explaining it.

Errors, and what they mean

Checking an endpoint is alive

Every endpoint answers a plain GET with a description of itself, so you can confirm what is deployed without reading any documentation — including the one you are reading now.

# The speech endpoint describes its own contract
curl -s https://zyraxonai.lovable.app/api/public/stt | jq '{modes, containers, limits}'

# The models the pool is serving right now
curl -s https://zyraxonai.lovable.app/api/public/zyraxon/v1/models | jq -r '.data[] | .id'

# Which voices exist, and how many
curl -s https://zyraxonai.lovable.app/api/public/tts/voices | jq '{count, languages: (.languages|length)}'