API Reference

158 models · 20 providers · OpenAI-compatible

Auth Scoped Keys Chat Embeddings Images Image Edits Videos Entities SPARQL Web Search Extract Scrape SEO Keywords Models Services Status Health Inference Quick Start

Authentication

All endpoints (except /health) require a Bearer token in the Authorization header.

Authorization: Bearer YOUR_API_KEY

Tip: Set your base URL to https://4ort.io/v1 and pass your API key — works with any OpenAI-compatible SDK or tool. Bare /chat/completions, /embeddings, etc. (without /v1/) also work.

Rate limits: Round-robin distributes across 18 locked free models (mostly NVIDIA NIM at 38 RPM each = ~684 RPM combined headroom). Per-model RPM tracked proactively — no wasted 429 attempts. Sustained 10-15 req/sec is comfortable; bursts higher work but increase tail latency.

Scoped API Keys

Alongside the admin key, the router issues named lane-restricted keys for individual consumers. A scoped key reaches exactly its lane — nothing else.

What a scoped key can do: depends on its grants. Chat keys reach POST /v1/chat/completions + GET /v1/models; serving-lane keys (cluster, geo, assertion) reach exactly their lane's endpoint — see the three cards below.

What is always denied: every other endpoint (403), models outside the key's provider allowlist (403), the AI Council (403), OpenRouter passthrough for unknown ids (403), and the configured fallback model if it sits outside the lane. With requireExplicitModel, "model": "auto" is denied too (403) — callers must name a model from their list.

SCOPED_API_KEYS=[{"name":"mint-verifier","key":"sk-getfree-mint-…","allowedProviders":["nvidia"],"freeModelsOnly":true,"requireExplicitModel":true,"dailyRequestLimit":6000}]

`name`	Label used in logs and analytics tagging (requests show as `name:X-App-Name`).
`key`	The bearer token. Generated on the router host, stored only in `/opt/getfree/.env` and on the consumer box — never in git, messages, or logs.
`allowedProviders`	Provider ids the key may route to (e.g. `nvidia`, `openrouter`). Empty list is a config error.
`freeModelsOnly`	Additionally require zero-priced models — the zero-cost guarantee.
`requireExplicitModel`	Deny auto-routing; every request must name a model id.
`dailyRequestLimit`	Server-side request cap per UTC day for this key (0 = unlimited). Exceeding it returns 429 with `api_key_daily_limit_reached`. Every capped-key response carries `X-RateLimit-Limit/-Remaining/-Reset` (fixed window, resets at UTC midnight); the 429 also carries `Retry-After` — honor it instead of retrying blind.

Issuing one for a new consumer: add an entry to SCOPED_API_KEYS in the router's .env (generate the key on the router host), then docker compose up -d proxy to apply. Hand the value to the consumer over a private channel, never in plaintext transit. The scoped GET /v1/models lists exactly the key's lane.

POST/v1/cluster/summary

Fusion cluster summarisation with coherence scoring — a local CPU model behind the gateway (ADR-0020). Batch variant: /v1/cluster/summary/batch with {"items": [...]}. Requires a scoped key with the cluster grant.

`cluster_id`	Stable id for the cluster (used in your own caching/dedup).
`headlines`	Array of headline strings — the cluster's contents.
`member_entities`	Optional per-member entity lists; enables the `entity_split` report (reported, never votes on `coherent`).

Response: {summary, coherent, coherence: {version, subject_overlap, uncorroborated_pairs, …}, ungrounded, truncated}. Coherence v2's stricter gate flags more incoherent clusters by design. Rate limit: per-key daily cap with X-RateLimit-* + Retry-After.

POST/v1/geo/generate

GEO draft generation (geo-v5 — current checkpoint, 6,007 rows, local INT8): page-title, H1 and slug transforms over the GEO envelope. Batch variant: /v1/geo/generate/batch. Requires a scoped key with the geo grant.

`task`	One of `title` · `h1` · `slug`. Required.
`entity`	The page's primary entity. Required.
`triples`	Verified fact triples grounding the generation. Required.
`brand / brand_qid / mode / page_type / keywords`	Optional envelope fields.

⚠️ Tier wall (enforced server-side, by design): only the draft tier is served. meta_description is HOLD-PRODUCTION — fabrication measured at 12% on v3 and the hold carries on v5, so the route refuses it with 403 geo_tier_held no matter how the caller is configured. faq_question/faq_answer are untrained on this model and return 422 geo_tier_untrained. The hosted teacher keeps the publish button for the held tier.

POST/v1/assertion

Assertion scoring (xnli, local INT8 model): is a triple asserted by any of its candidate sentences? Multi-sentence max-wins; winning_sentence_idx traces each decision to its text. Batch variant: /v1/assertion/batch. Requires a scoped key with the assertion grant.

`sentences`	Array of candidate sentences. Required.
`head / relation / tail`	The triple to check. All three required.
`lang`	Optional language tag, passed through.

Response: {p_asserted, winning_sentence_idx, decision: "accept"|"reject"|"abstain", per_sentence, hypothesis, latency_ms}. Thresholds (accept ≥ 0.97 / reject < 0.30) are consumer-owned and frozen — the route is a thin authenticated proxy and never adjusts them.

POST/v1/chat/completions

OpenAI-compatible chat completions. Supports streaming, tool calling, vision, and structured output. 130+ chat models auto-selected by priority, or specify one explicitly. Models not in the registry are passed through to OpenRouter automatically.

Parameter	Type	Description
messages	array	Array of message objects with `role` and `content`. Required.
model	string	Model ID (e.g. `google/gemini-2.5-flash`). Optional — auto-selects best available model if omitted.
stream	boolean	Enable SSE streaming. Default: false
max_tokens	integer	Maximum tokens to generate. Optional
temperature	number	Sampling temperature 0-2. Optional
tools	array	Tool/function definitions for tool calling. Optional
response_format	object	Structured output format, e.g. `{"type":"json_object"}`. Optional

// Request
{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is 2+2?"}
  ],
  "model": "google/gemini-2.5-flash",
  "stream": false,
  "max_tokens": 1024
}

// Response
{
  "id": "chatcmpl-abc123",
  "object": "chat.completion",
  "model": "google/gemini-2.5-flash",
  "choices": [{
    "index": 0,
    "message": {
      "role": "assistant",
      "content": "2 + 2 = 4."
    },
    "finish_reason": "stop"
  }],
  "usage": {"prompt_tokens": 25, "completion_tokens": 8, "total_tokens": 33}
}

Auto-routing: Omit the model field and the proxy will automatically select the highest-priority healthy model. If a model fails, the request is automatically retried on the next best model.

Routing Headers (optional)

Header	Values	Description
X-Routing	round-robin, priority	Load-balancing strategy. `round-robin` distributes across models; `priority` always picks the top model first. Default: server config
X-Parallel	1-5	Number of models to try simultaneously. `1` = sequential (saves quota), `2-3` = fast failover (uses more quota). Default: server config
X-No-Fallback	true	When set with an explicit model, disables all fallback behavior. If the requested model fails, returns the error directly instead of trying other models. Useful for premium/paid models where you want consistent quality. Default: off
X-App-Name	string	Tag requests with your app name for usage tracking and analytics. Optional

// Example: fast mode — round-robin + 3 parallel attempts
X-Routing: round-robin
X-Parallel: 3

// Example: quota-saver — priority routing, no parallelism
X-Routing: priority
X-Parallel: 1

// Example: premium model — no fallback to free models
X-No-Fallback: true

Throughput Benchmarks

Real-world load test results with round-robin routing and per-model RPM limiting (April 2026).

Concurrency	Requests	Success	p50 Latency	p90 Latency
15	100	99%	1.6s	90s
20	200	79.5%	963ms	37s

Recommendation: Stay under 15 concurrent requests for ~99% success and snappy p50. Sub-second median latency on small/medium models. Big models (Mistral Large 675B, DeepSeek 671B) can take 60-90s — set client timeout to 120s+. Big-model timeouts mostly happen at high concurrency.

POST/v1/embeddings

OpenAI-compatible text embeddings across 5 models. Auto-selects highest-priority available model — no fallback (embedding dimensions differ between providers).

Parameter	Type	Description
input	string \| array	Text or array of texts to embed. Required.
model	string	Model ID. Optional — auto-selects highest priority available.

Available Models

Model ID	Dim	Notes
4ort/qwen3-embedding-4b	2000	Self-hosted Qwen3-Embedding-4B (truncated from 2560 for pgvector HNSW). Highest priority.
gemini/gemini-embedding-001	3072	Google Gemini embeddings. Free tier.
openai/text-embedding-3-large	3072	OpenAI large embeddings.
openai/text-embedding-3-small	1536	OpenAI small embeddings.
siliconflow/BAAI/bge-m3	1024	BAAI BGE-M3 multilingual.

// Request
{
  "input": "The quick brown fox jumps over the lazy dog",
  "model": "4ort/qwen3-embedding-4b"
}

// Response
{
  "object": "list",
  "data": [{
    "object": "embedding",
    "embedding": [0.0023, -0.0091, 0.0152, ...],
    "index": 0
  }],
  "model": "4ort/qwen3-embedding-4b",
  "usage": {"prompt_tokens": 9, "total_tokens": 9}
}

Batch input: Pass an array of strings to input to embed multiple texts in a single request — far more efficient than one request per text.

No auto-fallback: Different embedding models have different dimensions (1024-3072), so a fallback would silently change the vector shape. Pin a model explicitly if you're storing vectors in a database.

POST/v1/images/generations

OpenAI-compatible image generation. 23 image models across Gemini (Nano Banana 2/Pro), OpenAI (GPT Image 1.5, DALL-E 3), x.ai (Grok Imagine), and SiliconFlow (FLUX, Qwen-Image, Stable Diffusion, Kolors). Auto-selects best model or specify one.

Parameter	Type	Description
prompt	string	Text description of the image to generate. Required.
model	string	Model ID (e.g. `gemini/gemini-3.1-flash-image-preview`). Optional — auto-selects best.
size	string	Image size, e.g. `1024x1024`, `1536x1024`. Default: 1024x1024
n	integer	Number of images to generate (1-4). Default: 1
quality	string	Quality level: `standard` or `hd`. Optional, model-dependent

// Request
{
  "prompt": "A futuristic city skyline at sunset, cyberpunk style",
  "model": "gemini/gemini-3.1-flash-image-preview",
  "size": "1024x1024",
  "n": 1
}

// Response
{
  "created": 1740600000,
  "data": [
    {
      "b64_json": "iVBORw0KGgo...",
      // or "url": "https://..." depending on model
    }
  ]
}

Response format varies by provider: Gemini and OpenAI return base64 in b64_json (heavy — ~100KB+ per image). x.ai and SiliconFlow return temporary url. All wrapped in OpenAI's standard { data: [...] } shape. Tip: if you're feeding the response back into an LLM context window, prefer providers that return URLs.

POST/v1/images/edits

OpenAI-compatible image editing — edit an existing image with a text prompt. Currently routed through x.ai Grok Imagine. Supports single or multi-image inputs.

Parameter	Type	Description
prompt	string	Edit instruction (e.g. "Make the sky red"). Required.
image	string \| array	URL or base64 data URI of the source image. Required (or use `images`).
images	array	Array of URLs or base64 data URIs for multi-image edits. Alternative to `image`.
model	string	Model ID. Optional — auto-selects.
n	integer	Number of variations to generate. Default: 1
size	string	Output size. Optional

// Request
{
  "prompt": "Add a sunset glow and dramatic clouds",
  "image": "https://example.com/photo.jpg",
  "model": "xai/grok-imagine-image"
}

POST/v1/videos/generations

Async video generation via 6 models across x.ai and fal.ai. Returns a request_id — poll GET /v1/videos/:requestId for the result.

Parameter	Type	Description
prompt	string	Text description of the video to generate. Required.
model	string	Model ID — required. Available: `xai/grok-imagine-video`, `fal/minimax-hailuo`, `fal/kling-v2.1`, `fal/wan-2.6`, `fal/hunyuan-video`, `fal/pika-2.2`
duration	integer	Video duration in seconds (model-dependent max: 5-15s). Optional
aspect_ratio	string	Aspect ratio: `1:1`, `16:9`, `9:16`, `4:3`, etc. Optional, varies by model
resolution	string	`480p`, `720p`, or `1080p`. Optional, varies by model
image_url	string	URL or base64 data URI for image-to-video generation. Optional (xAI only)
source_video_url	string	URL of an existing video to extend (continuation). New duration is appended to the source. Source must be ≤15s. Optional (xAI only)

// Step 1: Create video generation job
POST /v1/videos/generations

{
  "prompt": "A rocket launching from a tropical island at sunset",
  "model": "xai/grok-imagine-video",
  "duration": 10,
  "aspect_ratio": "16:9",
  "resolution": "720p"
}

// Response — job created (async)
{
  "request_id": "vg_abc123def456",
  "status": "in_progress"
}

GET/v1/videos/:requestId

Poll for video generation status. When status is "completed", the response includes the video URL.

// Step 2: Poll for result
GET /v1/videos/vg_abc123def456

// Response — completed
{
  "status": "completed",
  "video": {
    "url": "https://...",
    "duration": 10
  }
}

Async workflow: Video generation takes 30-180 seconds. Poll the status endpoint every 5-10 seconds until status is "done" or "failed". Pricing: $0.04-$0.08/second depending on model.

Continuation chain: To build a longer narrative, generate a base clip → grab its video.url from the status response → POST a new request with source_video_url set to it. Repeat to chain. xAI keeps style/character consistent from the source clip's last frame. Source must be ≤15 seconds.

Request ID prefix: Returned IDs are prefixed with the provider (e.g. xai_abc123) so the same status endpoint works regardless of which provider generated it.

POST/v1/entities/search

Unified entity search across 4 knowledge graph providers. Each provider returns standardized Entity objects with provider-specific metadata for SEO, academic research, and identity resolution.

Parameter	Type	Description
query	string	Search query. Required.
provider	string	One of: `wikidata`, `openalex`, `crossref`, `orcid`. Default: wikidata
limit	integer	Max entities to return (1-20). Default: 10
language	string	Language code (Wikidata only). Default: en

Providers

Provider	Best For	Metadata Fields
wikidata	General entities, QIDs, cross-linking	QID, types
openalex	Academic works, authors, institutions	citedByCount, worksCount, externalId
crossref	Scholarly works, DOI resolution	authors, publisher, journal, citationCount, ISSN, license
orcid	Researcher profiles, identity resolution	keywords, externalIds, urls

// Request
{
  "query": "Python programming",
  "provider": "wikidata",
  "limit": 5
}

// Response
{
  "query": "Python programming",
  "provider": "wikidata",
  "entities": [
    {
      "id": "Q28865",
      "name": "Python",
      "types": [],
      "description": "general-purpose programming language",
      "url": "https://www.wikidata.org/wiki/Q28865",
      "image": null,
      "score": null,
      "metadata": { /* provider-specific fields */ }
    }
  ]
}

Cross-linking: Wikidata QIDs are the universal identifier — many entities have ORCID iDs (P496), DOIs (P356), and OpenAlex IDs (P10283) attached. Use SPARQL to traverse these relationships.

ORCID note: ORCID search is a two-step lookup (search + per-profile fetch) so it's slower but returns much richer person data than the others.

POST/v1/entities/sparql

Raw SPARQL queries against Wikidata's graph database. Returns unprocessed bindings for maximum flexibility — relationship traversal, complex filters, and cross-linking entities to ORCID, DOI, OpenAlex.

Parameter	Type	Description
query	string	Valid SPARQL query against the Wikidata endpoint. Required.

Useful Properties

Property	Meaning
wdt:P31	instance of (e.g. `wd:Q5` = human)
wdt:P279	subclass of
wdt:P361	part of
wdt:P496	ORCID iD — links to ORCID provider
wdt:P356	DOI — links to CrossRef provider
wdt:P10283	OpenAlex ID — links to OpenAlex provider

// Request — Find Nobel Prize laureates
{
  "query": "SELECT ?item ?label WHERE { ?item wdt:P31 wd:Q5 . ?item wdt:P166 wd:Q35637 . ?item rdfs:label ?label . FILTER(lang(?label)='en') } LIMIT 5"
}

// Response
{
  "query": "SELECT ?item ...",
  "provider": "wikidata-sparql",
  "results": [
    { "item": { "value": "http://www.wikidata.org/entity/Q937" }, "label": { "value": "Albert Einstein" } }
  ]
}

POST/v1/search

Web search via self-hosted SearXNG metasearch. Privacy-respecting, aggregates DuckDuckGo, Brave, Wikipedia, Wikidata, ArXiv, and more.

Parameter	Type	Description
query	string	Search query. Required.
categories	string \| array	Categories: `general`, `images`, `news`, `science`, `files`, `music`, `videos`, `it`, `social_media`. Optional
engines	string \| array	Specific engines: `duckduckgo`, `brave`, `wikipedia`, `wikidata`. Optional
language	string	Language code (e.g. `en`). Default: en
pageno	number	Page number for pagination. Default: 1
time_range	string	`day`, `week`, `month`, `year`. Optional
safesearch	number	0 = off, 1 = moderate, 2 = strict. Default: 0

// Response
{
  "query": "typescript generics",
  "number_of_results": 5,
  "results": [
    {
      "title": "TypeScript: Documentation - Generics",
      "url": "https://www.typescriptlang.org/docs/handbook/2/generics.html",
      "content": "Generics provide a way to make components work with any data type...",
      "engine": "brave",
      "category": "general",
      "score": 1.0,
      "publishedDate": null,
      "thumbnail": null
    }
  ],
  "suggestions": [],
  "infoboxes": []
}

POST/v1/extract

Extract clean article content from any URL using Mozilla Readability. Returns title, author, date, and markdown body. Lightweight alternative to Firecrawl for article extraction.

Parameter	Type	Description
url	string	URL to extract content from. Required.
format	string	`markdown` or `text`. Default: markdown

// Response
{
  "title": "Article Title",
  "author": "Author Name",
  "date": "2026-04-17",
  "content": "Clean article text in markdown...",
  "excerpt": "First paragraph summary...",
  "url": "https://example.com/article"
}

POST/v1/scrape

Extract structured data from any URL using CSS selectors. Server-side DOM parsing via linkedom — works on any HTML page.

Parameter	Type	Description
url	string	URL to scrape. Required.
selectors	object	Key-value pairs of CSS selectors. Each value extracts text from matching element(s). Required.

// Request
{
  "url": "https://example.com",
  "selectors": {
    "title": "h1",
    "description": "meta[name='description']",
    "links": "a[href]"
  }
}

// Response
{
  "url": "https://example.com",
  "data": {
    "title": "Example Domain",
    "description": "An example website",
    "links": ["More information..."]
  }
}

SEO Keyword Research

DataForSEO-powered SEO keyword endpoints — search volume, difficulty, suggestions, intent classification, and AI-search keyword research.

POST/v1/seo/keywords/volume

Get monthly search volume, CPC, competition for up to 1,000 keywords from Google Ads.

// Request
{
  "keywords": ["typescript", "javascript"],
  "location": "United States",    // or location code (e.g. 2840)
  "language": "en"
}

POST/v1/seo/keywords/research

Full keyword research data — volume, CPC, search intent, SERP info — for up to 700 keywords. Requires location + language.

POST/v1/seo/keywords/suggestions

Generate related keyword suggestions from a single seed keyword.

// Request
{
  "keyword": "running shoes",
  "location": "United States",
  "language": "en",
  "limit": 100,
  "include_seed": true
}

POST/v1/seo/keywords/difficulty

SEO difficulty scores (0-100) showing how hard it is to rank for each keyword. Up to 1,000 keywords. Requires location + language.

POST/v1/seo/keywords/intent

Classify search intent (informational, navigational, commercial, transactional) for each keyword. Up to 1,000. Requires language.

POST/v1/seo/keywords/ai-volume

AI search volume (queries to ChatGPT, Perplexity, Claude, etc.) for up to 1,000 keywords. Useful for tracking AI-driven traffic. Requires location + language.

Location: Pass either a country name (e.g. "United States") or a DataForSEO location code (e.g. 2840).

Pricing: SEO endpoints are not free — billed via the underlying DataForSEO account. Other endpoints on 4ort.io are free or use the proxy's free-tier providers.

GET/v1/models

List all available models — chat, image, video, embedding, and entity-search. Each entry has a type field and a pricing object. 0 = free.

// Response (truncated)
{
  "object": "list",
  "data": [
    {
      "id": "nvidia/qwen/qwen3-coder-480b-a35b-instruct",
      "object": "model",
      "type": "chat",
      "owned_by": "nvidia",
      "pricing": {"input": 0, "output": 0},
      "capabilities": {"tools": true, "vision": false, "reasoning": false}
    },
    {
      "id": "gemini/gemini-3.1-flash-image-preview",
      "type": "image",
      "pricing": {"perImage": 0.067},
      "supported_sizes": ["1024x1024", "1536x1024", "1024x1536"]
    },
    {
      "id": "xai/grok-imagine-video",
      "type": "video",
      "pricing": {"perSecond": 0.05},
      "async": true,
      "endpoint": "/v1/videos/generations"
    },
    {
      "id": "4ort/qwen3-embedding-4b",
      "type": "embedding",
      "dimensions": 2000,
      "pricing": {"input": 0}
    }
  ]
}

Pricing fields: Chat: pricing.input/output in $/million tokens. Image: pricing.perImage. Video: pricing.perSecond. Embedding: pricing.input. Entity search: pricing.perRequest.

GET/v1/services

Programmatic service discovery — returns all enabled endpoints, their parameter schemas, and provider lists. Useful for agents that need to discover capabilities at runtime.

// Response (excerpted)
{
  "services": {
    "chat": { "endpoint": "POST /v1/chat/completions", "params": { /* ... */ } },
    "images": { "endpoint": "POST /v1/images/generations", "providers": ["gemini/...", "openai/..."] },
    "videos": { "endpoints": ["POST /v1/videos/generations", "GET /v1/videos/:requestId"] },
    "entity_search": { "providers": ["wikidata", "openalex", "crossref", "orcid"] },
    "sparql": { "endpoint": "POST /v1/entities/sparql" },
    "embeddings": { "providers": ["4ort/qwen3-embedding-4b", "gemini/..."] },
    "web_search": { "endpoint": "POST /v1/search" },
    "seo_keywords": { "endpoints": ["POST /v1/seo/keywords/volume", /* ... */] }
  }
}

GET/v1/status

Live model rankings, health, rate-limit usage, and routing breakdown. Updated continuously from real request data — useful for monitoring which models are healthy and available right now.

// Response (excerpted)
{
  "totalModels": 158,
  "imageModels": 23,
  "videoModels": 6,
  "embeddingModels": 5,
  "entitySearchModels": 4,
  "providers": {"nvidia": 35, "openrouter": 85, /* ... */},
  "rateLimits": {/* per-provider daily/RPM usage */},
  "models": [/* ranked by effective priority + health */]
}

GET/health

Public health check. No authentication required. Returns server status and model counts.

// Response
{"status": "ok", "uptime": 86400, "totalModels": 158}

GET/v1/inference/health

Status of the self-hosted inference server (where 4ort/* models run). Returns CPU, memory, per-model status, queue depth, and request stats. Cached for 3 seconds — safe to poll often.

// Response (excerpted)
{
  "available": true,
  "status": "ok",                         // ok | degraded | down
  "uptime_seconds": 12345,
  "system": {
    "cpu_percent": 42.3,
    "memory_used_gb": 28.4,
    "memory_total_gb": 256,
    "load_avg": [3.2, 3.5, 3.8]
  },
  "models": {
    "fast": { "name": "Qwen3.5-0.8B", "status": "idle", "queue_depth": 0, "requests_ok": 1542, "avg_latency_ms": 342 },
    "quality": { "name": "GLM-4.7-Flash-REAP-23B", "status": "processing", /* ... */ },
    "embedding": { "name": "Qwen3-Embedding-4B", "status": "idle", /* ... */ }
  }
}

Quick Start

from openai import OpenAI
import requests, time

client = OpenAI(
    base_url="https://4ort.io/v1",
    api_key="YOUR_API_KEY",
)

# Chat completion
response = client.chat.completions.create(
    model="google/gemini-2.5-flash",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

# Image generation
image = client.images.generate(
    model="gemini/gemini-3.1-flash-image-preview",
    prompt="A cat astronaut floating in space",
    size="1024x1024",
)
print(image.data[0].b64_json[:50])

# Embeddings
emb = client.embeddings.create(
    model="4ort/qwen3-embedding-4b",
    input=["hello world", "goodbye world"],
)
print(f"vec dim: {len(emb.data[0].embedding)}")

# Video generation (async)
headers = {"Authorization": f"Bearer {client.api_key}",
           "Content-Type": "application/json"}
job = requests.post("https://4ort.io/v1/videos/generations",
    json={"prompt": "A rocket launch at sunset",
          "model": "xai/grok-imagine-video"},
    headers=headers).json()
while True:
    status = requests.get(f"https://4ort.io/v1/videos/{job['request_id']}",
        headers=headers).json()
    if status["status"] in ("done", "failed"): break
    time.sleep(5)
print(status)

# Entity search (Wikidata)
entities = requests.post("https://4ort.io/v1/entities/search",
    json={"query": "Albert Einstein", "provider": "wikidata"},
    headers=headers).json()
for e in entities["entities"]:
    print(f"{e['name']}: {e['description']}")

# Web search
search = requests.post("https://4ort.io/v1/search",
    json={"query": "latest typescript features", "engines": "brave,duckduckgo"},
    headers=headers).json()
for r in search["results"][:3]:
    print(r["title"], r["url"])

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://4ort.io/v1",
  apiKey: "YOUR_API_KEY",
});

// Chat completion
const response = await client.chat.completions.create({
  model: "google/gemini-2.5-flash",
  messages: [{role: "user", content: "Hello!"}],
});
console.log(response.choices[0].message.content);

// Image generation
const image = await client.images.generate({
  model: "gemini/gemini-3.1-flash-image-preview",
  prompt: "A cat astronaut floating in space",
  size: "1024x1024",
});
console.log(image.data[0].b64_json?.slice(0, 50));

// Embeddings
const emb = await client.embeddings.create({
  model: "4ort/qwen3-embedding-4b",
  input: ["hello world", "goodbye world"],
});
console.log(`vec dim: ${emb.data[0].embedding.length}`);

// Video generation (async)
const headers = {
  "Authorization": `Bearer ${client.apiKey}`,
  "Content-Type": "application/json",
};
const job = await fetch("https://4ort.io/v1/videos/generations", {
  method: "POST", headers,
  body: JSON.stringify({
    prompt: "A rocket launch at sunset",
    model: "xai/grok-imagine-video",
  }),
}).then(r => r.json());
let status;
do {
  await new Promise(r => setTimeout(r, 5000));
  status = await fetch(`https://4ort.io/v1/videos/${job.request_id}`,
    {headers}).then(r => r.json());
} while (!["done", "failed"].includes(status.status));
console.log(status);

// Entity search (Wikidata)
const entities = await fetch("https://4ort.io/v1/entities/search", {
  method: "POST", headers,
  body: JSON.stringify({query: "Albert Einstein", provider: "wikidata"}),
}).then(r => r.json());
entities.entities.forEach(e =>
  console.log(`${e.name}: ${e.description}`));

# Chat completion
curl https://4ort.io/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Hello!"}],
    "model": "google/gemini-2.5-flash"
  }'

# Image generation
curl https://4ort.io/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A cat astronaut floating in space",
    "model": "gemini/gemini-3.1-flash-image-preview"
  }'

# Embeddings
curl https://4ort.io/v1/embeddings \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "hello world", "model": "4ort/qwen3-embedding-4b"}'

# Video generation — Step 1: Create job
curl https://4ort.io/v1/videos/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "A rocket launch at sunset", "model": "xai/grok-imagine-video"}'
# → {"request_id": "xai_abc123", "model": "xai/grok-imagine-video"}

# Video generation — Step 2: Poll for result
curl https://4ort.io/v1/videos/xai_abc123 \
  -H "Authorization: Bearer YOUR_API_KEY"
# → {"status": "done", "video": {"url": "...", "duration": 10}}

# Web search
curl https://4ort.io/v1/search \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "typescript generics", "engines": "brave,duckduckgo"}'

# Article extraction
curl https://4ort.io/v1/extract \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/article"}'

# Entity search — Wikidata
curl https://4ort.io/v1/entities/search \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "Python programming", "provider": "wikidata"}'

Home Dashboard