ZYRAXONAI Platform

Provider

Ten models. One endpoint. Yours.

ZYRAXON serves its own line of ten branded models over a single OpenAI-compatible API. Point any OpenAI client at the endpoint below, pick a ZYRAXON model, and build. No upstream keys, no provider switching, no rate-limit juggling.

QuickstartUse it anywhereIntegration guideSee the modelsOpenAI-compatibleStreamingMultimodal

Endpoint

Two URLs, that is the whole integration

The models list is what a client reads to discover us. The chat URL is what it calls to use us.

GETModel catalogue
https://zyraxonai.lovable.app/api/public/zyraxon/v1/models

Returns the ten ZYRAXON models in the standard OpenAI { object: "list", data: [...] } shape, so existing clients read it without special handling.

POSTChat completions
https://zyraxonai.lovable.app/api/public/zyraxon/v1/chat/completions

OpenAI chat-completions semantics, including server-sent-event streaming when "stream": true. Responses name the ZYRAXON model you asked for.

The line

Ten ZYRAXON models

Each model is built for a different job. Capabilities below are read live from the serving pool, so what you see is what is answering right now.

ZYRAXON Apex

Flagship reasoning at a million tokens

Deep reasoning · 1M context · 131K output

zyraxon-apeximagevideoPDF

ZYRAXON Matrix

Sees images, reasons at scale

Vision + text · 1M context · deep reasoning

zyraxon-matriximage

ZYRAXON Pulse

Built for speed on long inputs

Low latency · 262K context · 262K output

zyraxon-pulsetext

ZYRAXON Titan

The long-output workhorse

1M context · 512K output · image + video

zyraxon-titanimagevideo

ZYRAXON Neural

Balanced general intelligence

Reasoning · 200K context · 32K output

zyraxon-neuraltext

ZYRAXON Codex

Writes and repairs code

Code + vision · 200K context

zyraxon-codeximage

ZYRAXON Visionary

Reads images, video and documents

Image + video + PDF · 200K context

zyraxon-visionaryimagevideoPDF

ZYRAXON Cyber

Every input modality, one model

Text + image + audio + video · 200K context

zyraxon-cyberimagevideoaudio

ZYRAXON Quantum

Million-token document reasoning

1M context · vision · 131K output

zyraxon-quantumimagePDF

ZYRAXON Genesis

Fresh from the lab

Latest experimental reasoning · every modality

zyraxon-genesisimagevideoaudioPDF

Automatic rotation

Each ZYRAXON model is bound to a preferred model underneath, with every other model in the pool behind it as a fallback. If one is retired or rate-limited, the same ZYRAXON model keeps answering from the next one — so a Rotated badge means it is covering, not down. You always call the same name; we handle the swap.

Multimodal

Send images, audio, video and documents

Chat completions accept the OpenAI content-parts array, so one message can mix text with media. When a turn carries media, it is routed to a model that can actually see it — a video turn is never sent to a text-only model and made to fail.

Vision — image + textjson
{
  "model": "zyraxon-matrix",
  "messages": [{
    "role": "user",
    "content": [
      { "type": "text", "text": "What is in this chart?" },
      { "type": "image_url",
        "image_url": { "url": "data:image/png;base64,..." } }
    ]
  }]
}
Audio, video, documentsjson
{
  "model": "zyraxon-cyber",
  "messages": [{
    "role": "user",
    "content": [
      { "type": "text", "text": "Transcribe and summarise" },
      { "type": "input_audio",
        "input_audio": { "data": "<base64>", "format": "wav" } },
      { "type": "video_url",
        "video_url": { "url": "data:video/mp4;base64,..." } },
      { "type": "file",
        "file": { "filename": "report.pdf", "file_data": "data:application/pdf;base64,..." } }
    ]
  }]
}

Big video just works

A long recording is compressed and sliced in the browser before upload, so a multi-gigabyte file becomes a handful of model-sized parts instead of a failed request.

Capability-aware routing

The modalities each model accepts are read live from the serving pool, and a media turn is ordered onto the models that can handle it.

Told what was used

A media response carries X-ZYRAXON-Input-Modalities, so you can tell a real multimodal answer from a text-only one.

Quickstart

Point your client at us

Any OpenAI-compatible client works. Two examples below.

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

const client = new OpenAI({
  baseURL: "https://zyraxonai.lovable.app/api/public/zyraxon/v1",
  apiKey: "not-needed",
});

const stream = await client.chat.completions.create({
  model: "zyraxon-codex",
  messages: [{ role: "user", content: "Build a landing page" }],
  stream: true,
});

Response shape

Standard chat-completion chunks. The model field carries the ZYRAXON name you requested, so your logs and billing stay yours.

data: {"id":"zyraxon-...","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"Hello"},"finish_reason":null}],"model":"ZYRAXON Apex"}

Anywhere

Use it from anything

Termux on a phone, OpenHands, an editor, a mobile app, an SDK, a workflow runner — they all connect the same way. Base URL https://zyraxonai.lovable.app/api/public/zyraxon/v1, model zyraxon-codex, any non-empty API key.

Base URL

https://zyraxonai.lovable.app/api/public/zyraxon/v1

Default model

zyraxon-codex

API key

zyraxon
Read the integration guideMachine-readable catalogue
Terminal & Mobile

Termux (Android)

Run ZYRAXON from the terminal on an Android phone. The setup script installs curl and jq, stores your settings, and adds a zyraxon command that opens an interactive chat box — type a message, get a streamed reply, switch models, and turn on tool calling so the agent can read, write and run things on the phone.

AndroidChat boxAny modelTool calls

Kali Linux (desktop and NetHunter)

Kali is a Debian-family distro with a shell, so the provider works exactly as it does on any Linux box — and Kali NetHunter puts the same client on an Android phone inside its chroot. One zyraxon-ask script gives you a streaming prompt with your environment already wired.

KaliNetHunterDebianShell

Mobile apps (React Native, Flutter, Android, iOS)

A mobile app does not need an OpenAI SDK. The endpoint is plain HTTPS with CORS open to *, so the platform HTTP client is enough — and the media parts are the same shapes the web client already sends.

React NativeFlutterKotlinSwift
Agents & Editors

OpenHands

OpenHands takes a custom provider through a small form. Copy the four fields below and it uses the ZYRAXON line like any other OpenAI-compatible model.

Custom providerBase URL + API key

Cursor, Continue, Cline, Aider and any OpenAI-compatible app

Anything with an *OpenAI-compatible* or *custom base URL* setting connects with the same two values. This is the largest group of clients, and the one that needs no code at all.

Drop-inNo code

VS Code: Cline, Roo Code, Kilo Code and Copilot Chat

The VS Code agent extensions all expose an *OpenAI Compatible* provider, so they connect with the same base URL and key as everything else. Copilot Chat and extensions that read the OpenAI environment variables work too.

VS CodeClineRoo CodeCopilot

Zed

Zed takes a custom OpenAI-compatible URL in its settings, so the assistant and the inline panel both run on the ZYRAXON line.

ZedEditor

Ollama-style and LocalAI-style clients

Clients that expect a local model server on localhost can be pointed at us instead, because they speak the same OpenAI-compatible dialect.

Drop-inLocal server replacement

MCP servers and agent tool hosts

An MCP server or any agent framework can use ZYRAXON as its model and keep its own tool definitions — they travel in the request and come back as real tool calls.

MCPTool calls
SDKs & Frameworks

cURL

The shortest path to a working call. Two headers and a JSON body, and it works from any machine with curl — including a phone.

AnywhereNo install

OpenAI SDK (Node and Python)

The official SDKs work unchanged — only the base URL moves. This is the path to use when you want retries, typed responses and streaming without writing HTTP by hand.

TypeScriptPythonStreaming

Vercel AI SDK

Use the OpenAI-compatible provider from @ai-sdk/openai-compatible. This is the same transport the ZYRAXON builder gateway uses, so the shapes are already proven.

TypeScriptStreaming

LangChain

LangChain's OpenAI integrations accept a custom base URL, so the whole chain, agent and tooling surface is available over the ZYRAXON line.

PythonJavaScript

LiteLLM (and Open WebUI, LibreChat, Dify, Flowise)

LiteLLM refuses a model string without a provider prefix — that is the "LLM Provider NOT provided" error. Write the model as openai/zyraxon-codex and it uses the generic OpenAI-compatible handler. Every LiteLLM-based gateway (Open WebUI, LibreChat, Dify, Flowise, Langroid) connects the same way.

PythonProxyOpen WebUIGateways

Python shell, REPL and notebooks

Import one dependency-free file and the provider is a function call. Works in the plain REPL, IPython, Jupyter, Colab, a script or a cron job. Every model works — the list is read live, so nothing is pinned.

PythonREPLJupyterTool calls

PowerShell

Dot-source one script and the provider becomes four functions. Works in Windows PowerShell and PowerShell 7 on every platform, with functions for a plain answer, a streamed answer, and the tool-calling agent loop.

PowerShellWindowsTool calls

Bash and zsh

Source one file in the shell you already have open and the provider becomes shell functions. Composes with pipes, so a log, a diff or a command's output can be handed straight to the model.

BashzshPipesTool calls
Languages & Runtimes

Go

The standard library is enough — no third-party client required. This works from a CLI tool, a service or a Go worker.

Gonet/http

Rust

reqwest plus serde_json, which most Rust services already depend on. Streaming works the same way over the same URL.

Rustreqwest

PHP

Plain cURL, so it drops into any PHP app, Laravel job, Symfony command or WordPress plugin without a client library.

PHPcURL

Java, Kotlin and the JVM

The JDK's own HTTP client, so Spring Boot, a Gradle task, an Android app or a plain CLI can call the endpoint with no extra dependency.

JavaJVMSpring

.NET and C#

HttpClient directly, or the OpenAI .NET SDK with a custom endpoint. Either way the request shape is the same one every other client sends.

.NETC#
Bots & Home

Discord, Slack and chat bots

A bot is one HTTP call. Take the message, send it as a user turn, post the reply back — the same pattern works for Discord, Slack, Telegram or a chat widget on your own site.

DiscordSlackBots

Home Assistant, Node-RED and smart-home

Home automation hosts are HTTP clients with a templating layer, so a REST command or an HTTP request node is the whole integration.

Home AssistantNode-REDIoT
Automation & Platforms

n8n, Make and Zapier

No-code runners reach the endpoint through an HTTP request node. Point the node at the chat URL and map your previous step's output into the message content.

No codeHTTP node

Postman, Insomnia and other HTTP clients

For exploring the API by hand before writing code. Save the request, then swap the model id to see how another ZYRAXON model answers the same turn.

Manual testing

Voice input and speech output

Speech-to-text and text-to-speech are both endpoints. Record on the device, transcribe it, send the transcript as a normal chat turn — then turn any answer back into speech, in the language the answer was written in, streamed as MP3 as it is synthesised.

Speech to textText to speech322 voices75 languages

Speech

It can speak, in any language

Every answer can be heard, not just read — and in the language the answer was written in. Ask in Bengali and a Bengali voice reads the reply. The same engine behind Edge's Read Aloud: 322 neural voices across 75 languages, streamed as MP3 as it is written.

Hear an answerPOST /api/public/tts
curl -s https://zyraxonai.lovable.app/api/public/tts \
  -H "Content-Type: application/json" \
  -d '{"text":"Hello from ZYRAXON.","locale":"en"}' \
  -o answer.mp3
The same in Bengalilocale: bn
curl -s https://zyraxonai.lovable.app/api/public/tts \
  -H "Content-Type: application/json" \
  -d '{"text":"আমি একটি ZYRAXON মডেল।","locale":"bn"}' \
  -o bengali.mp3

# bn-BD-NabanitaNeural reads it — a Bengali voice,
# not an English one reading Bengali letters.

Every language

Name the language and the voice follows — Bengali, Hindi, Arabic, English and 71 more. The full list.

Streamed, not buffered

The MP3 starts arriving before the sentence is finished being synthesised, so playback begins almost immediately.

It speaks, not reads code

Code, links and Markdown are stripped first. A code block becomes “a code block, 12 lines” rather than being read character by character.

Speech in, speech out

Transcribe a recording with /api/public/stt, ask the model, then speak the answer. The whole loop is documented.

Read the speech guide

Why a rooted provider

One name you control

Stable names

Upstream models come and go. Your model value does not have to change when they do.

Capacity pooling

Every model in the pool stands behind every brand, so one model's limit is absorbed rather than passed to you.

Built to hold a flood

Each request runs in its own lane with its own fingerprint, so one caller's rate limit never stalls another. Under overload a request is shed fast with a 503 and Retry-After instead of queueing behind a stranger — the lanes already running stay healthy.

Drop-in

Same request and response shapes as OpenAI, so existing code works with a base-URL change.