Use it in your app

How to call it

Send your key in the Authorization header, and the picture (JPEG, PNG or WebP) as a form upload or as the body itself, at most 10 MB; or in JSON, as base64 ("image", at most 8 MB) or as a public https link to it ("image_url").

Ask our AI chips about a picture

curl https://ai.hypercluster.org/v1/vision/analyze \
  -H "Authorization: Bearer $HYPERAI_KEY" \
  -F [email protected] \
  -F tasks=objects,text

// Node 18+ (on your server, so the key stays private)
import fs from 'node:fs';

const form = new FormData();
form.append('image', new Blob([fs.readFileSync('gate.jpg')]), 'gate.jpg');
form.append('tasks', 'objects,text');
const res = await fetch('https://ai.hypercluster.org/v1/vision/analyze', {
  method: 'POST',
  headers: { Authorization: `Bearer ${process.env.HYPERAI_KEY}` },
  body: form,
});
const result = await res.json();
console.log(result.objects);

import os, requests

with open("gate.jpg", "rb") as f:
    r = requests.post(
        "https://ai.hypercluster.org/v1/vision/analyze",
        headers={"Authorization": f"Bearer {os.environ['HYPERAI_KEY']}"},
        files={"image": f},
        data={"tasks": "objects,text"},
    )
print(r.json()["objects"])

tasks is any of objects (things found, from 80 everyday kinds, each with a box), labels (what the picture shows, from 1,000 kinds), text (words and numbers, each with its four corners) and faces (how many, each with a box). Leave it out for all four.

What comes back

{
  "id": "4f1c…",
  "object": "vision.analysis",
  "image": { "width": 1920, "height": 1080 },
  "tasks": ["objects", "text"],
  "objects": [ { "label": "car", "score": 0.91, "box": [412, 380, 1210, 905] } ],
  "text": [ { "text": "WXY 1234", "score": 0.97, "box": [[640, 700], [880, 702], [879, 760], [639, 758]] } ],
  "models": { "yolov8n": "2026.10.1", "ppocrv4_det": "2026.10.1", "ppocrv4_rec": "2026.10.1" },
  "ms": 640,
  "cost": { "hypec": 0.45, "usd": 0.0045 },
  "balance": { "hypec": 99.55 }
}

Boxes are [left, top, right, bottom] in pixels of your picture, the right way up. Scores go from 0 to 1.

Run your own model

A model you keep in My models (a YOLO model you trained or uploaded: a .pt, or one exported to .onnx) runs on a member’s graphics card. It is paid for by the second it takes (at least one second, usually one to three), at the rate of the graphics card that runs it.

curl https://ai.hypercluster.org/v1/models/MODEL_ID/predict \
  -H "Authorization: Bearer $HYPERAI_KEY" \
  -F [email protected]

{
  "id": "9b2e…",
  "object": "model.prediction",
  "model": { "id": "MODEL_ID", "name": "Plate finder", "task": "detect" },
  "image": { "width": 1920, "height": 1080 },
  "objects": [ { "label": "plate", "class": 0, "score": 0.93, "box": [640, 700, 880, 760] } ],
  "gpu": "NVIDIA GeForce RTX 4070",
  "seconds": 1.8,
  "ms": 2410,
  "cost": { "hypec": 0.007146, "usd": 0.000071 },
  "balance": { "hypec": 99.54 }
}

A classification model answers with classes (its top five, each with a label and a score) instead of objects. The labels are the names you keep with the model in My models.

Chat with your fine-tuned model

A chat model you fine-tuned and keep in My models (its LoRA adapter) answers on a member’s graphics card, with the model it was trained on under it. It is paid for by the second it takes, as above. The first chat with it on a PC can take some minutes, while the PC fetches that model (not charged): then the answer is 202 with a poll address to ask again.

curl https://ai.hypercluster.org/v1/models/MODEL_ID/chat \
  -H "Authorization: Bearer $HYPERAI_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "messages": [ { "role": "user", "content": "Where is the office?" } ], "max_tokens": 256 }'

{
  "id": "5c1d…",
  "object": "model.chat",
  "status": "done",
  "reply": { "role": "assistant", "content": "Our office is in Kuala Lumpur." },
  "usage": { "tokensIn": 14, "tokensOut": 9 },
  "gpu": "NVIDIA GeForce RTX 3080 Laptop GPU",
  "seconds": 6.2,
  "cost": { "hypec": 0.03, "usd": 0.0003 },
  "balance": { "hypec": 99.51 }
}

What was said goes to the member’s PC and back; we keep the reply only until you collect it (10 minutes at most), and record only its tokens, time and cost.

Add a model

From a public https link straight to the file, for example a Hugging Face or GitHub download link: it is fetched, checked and kept encrypted, usually within a minute. Ask for the model until its status turns ready (or failed, with the reason). This needs a key you allowed to change models.

curl https://ai.hypercluster.org/v1/models \
  -H "Authorization: Bearer $HYPERAI_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "url": "https://huggingface.co/you/plates/resolve/main/best.pt", "name": "Plate finder" }'

A file on your computer goes up in pieces. Start with POST /v1/models and { "fileName": "best.pt", "bytes": 6200000 }: the answer gives the model's id and the largest piece. Then send each piece with PUT /v1/models/ID/chunk?offset=N, and finish with POST /v1/models/ID/done.

import os, requests

KEY = {"Authorization": f"Bearer {os.environ['HYPERAI_KEY']}"}
API = "https://ai.hypercluster.org/v1"
path = "best.pt"
r = requests.post(f"{API}/models", headers=KEY,
                  json={"fileName": path, "bytes": os.path.getsize(path), "name": "Plate finder"}).json()
model, size = r["model"]["id"], r["upload"]["chunkBytes"]
with open(path, "rb") as f:
    offset = 0
    while piece := f.read(size):
        requests.put(f"{API}/models/{model}/chunk", params={"offset": offset}, headers=KEY, data=piece).raise_for_status()
        offset += len(piece)
print(requests.post(f"{API}/models/{model}/done", headers=KEY).json()["status"])

A model is a YOLO model from Ultralytics, as a .pt or exported to .onnx (both run here), or another .onnx or PyTorch file (kept, but only YOLO models can run for now), at most 1 GB; you keep up to 1 GB of models free.

Also

• GET /v1/balance: your HYPEC balance.
• GET /v1/models: the models you keep, with their ids and where each stands (ready, preparing, uploading or failed), and whether it can run.
• PATCH /v1/models/ID: rename a model ({ "name": "…" }) or change its labels ({ "labels": [...] }); DELETE /v1/models/ID: delete it. Both need a key allowed to change models.
• GET /v1/openapi.json: all of it, described for tools that read OpenAPI.

Use it from an AI agent

HyperAI is also an MCP server, so an AI agent can use it as a tool: https://ai.hypercluster.org/mcp, with your API key in the Authorization header. Give the agent a key of its own with a monthly limit, so you can see what it spends and stop it at any time.

claude mcp add --transport http hyperai https://ai.hypercluster.org/mcp \
  --header "Authorization: Bearer $HYPERAI_KEY"

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "hyperai": {
      "url": "https://ai.hypercluster.org/mcp",
      "headers": { "Authorization": "Bearer hc_your_key" }
    }
  }
}

// .vscode/mcp.json
{
  "inputs": [{ "type": "promptString", "id": "hyperai-key", "description": "HyperAI API key", "password": true }],
  "servers": {
    "hyperai": {
      "type": "http",
      "url": "https://ai.hypercluster.org/mcp",
      "headers": { "Authorization": "Bearer ${input:hyperai-key}" }
    }
  }
}

Any other agent that connects to an MCP server over HTTP with a header works the same way. Agents that only connect by signing in (OAuth), such as the connectors in claude.ai and ChatGPT, can't use it yet.

What the agent can do:

What an agent did is in the list of calls above, marked as an AI agent’s, and in your statement.

When something is wrong

The answer is { "error": { "code", "message" } }, with these statuses:

Your pictures

Pictures sent to our AI chips go to machines HyperCluster runs. Pictures for your own model go to a member’s graphics card with your model, and are gone from it when the run ends; the PC may keep your model, still encrypted, for up to a week, so that its next run starts sooner. We don't keep the pictures: each call's time and cost are recorded for your statement, and, for our AI chips, a fingerprint of the picture, for checking their work.

HyperAI privacy · HyperAI terms