AIOZ Storage for AI and Data Pipelines: What's Real

AIOZ Storage for super intelligence (SI) work is S3-compatible object storage plus a vector database. You keep datasets, checkpoints and model outputs in normal buckets, and you store and search embeddings in vector buckets through the same S3-style API.

TL;DR:

• AIOZ Storage holds SI datasets, checkpoints and model outputs as plain S3 objects, so tools that already speak S3 need only a new endpoint and keys.
• The docs now ship a Vector Database with Python, JavaScript and Go guides, an MCP server for buckets, objects and vectors, and SI Memory with Record and Recall.
• Vectors are billed apart from files: $0.04 per GB uploaded, $0.05 per GB-month stored, plus a small fee per query.

Updated Oct 19, 2026: removed our old line saying AIOZ Storage had no vector database. The docs now cover a Vector Database, an MCP server and SI Memory, so this guide describes them and adds vector pricing.

What Does AIOZ Storage Offer for SI Workloads?

You get two kinds of storage under one account. Regular buckets hold objects such as training files, images, logs, checkpoints and exported results, while vector buckets hold embeddings in indexes that you query by similarity. Both run on AIOZ DePIN, which protects stored data with built-in Reed-Solomon erasure coding across its nodes, so you don't size or patch a separate cluster for either.

• Object storage: buckets, objects, encryption and scoped access grants. AIOZ Storage Explained covers the model, and AIOZ Storage Access Control covers the keys a pipeline would use.
• Vector Database: vector buckets and indexes that you reach with the AWS S3 Vectors SDKs. The concept page says existing S3 clients and SDKs "work out of the box".
• MCP server: lets an MCP-compatible assistant list and read objects, create upload links, manage vector buckets and run vector searches from plain prompts.
• Super Intelligence Memory: saves short statements with Record and returns the closest ones with Recall. You never build embeddings or indexes yourself. See our SI Memory guide.

If you are unsure which of the last two fits your agent, Vector Database vs Agent Memory on AIOZ Storage compares them side by side.

What the Docs Don't Cover Yet

AIOZ Storage has no managed training or inference service, and the docs don't describe a link to AIOZ AI's hosted models. The vector SDK guides also show the endpoint as a placeholder, so you take the vector API base URL from the SI Memory page instead. Everything else in this guide comes straight from the docs or from the third-party tools it names.

Why Object Storage Fits Datasets and Model Outputs

At the storage layer, training data, generated images, checkpoints and experiment outputs are all just objects. You write them once, read them many times, and find them by key instead of by a database schema, which is the exact shape S3 was built for.

It is also why almost every ML framework, from a short boto3 script to a managed training platform, already speaks S3. Pointing that tooling at https://s3.aiozstorage.network instead of AWS is a configuration change in the client. Our JavaScript, Go and Python guides show the same endpoint-and-keys swap. For the details of that role, see Using AIOZ Storage as a Dataset and Model-Output Backend.

Can You Checkpoint Agents to AIOZ Storage?

Yes, through LangGraph's own langgraph-checkpoint-aws package, which keeps agent state in DynamoDB and offloads checkpoints larger than 350 KB to S3. Its s3_offload_config takes a bucket_name and an optional endpoint_url, shown in the README with a localhost:4566 address.

You can put the AIOZ Storage endpoint in that same field, but keep in mind this is LangGraph's own flexibility, not an integration we built or tested. Offloading LangGraph Agent Checkpoints to AIOZ Storage walks through the setup and one credential-sharing catch.

How Do Agent Sandboxes Use Object Storage?

OpenAI's Agents SDK mounts external storage into an agent's workspace through a Manifest. The sandbox guide lists S3Mount, GCSMount, R2Mount, AzureBlobMount and BoxMount among its mount types. When your agent reads documents or saves generated files in that workspace, it is reading and writing objects through one of these mounts.

The SDK's API reference gives S3Mount an optional endpoint_url field, which is where a non-AWS endpoint would go. What Agent Sandboxes on Modal and E2B Store checks two sandbox providers directly.

Real Deployments on AIOZ Storage

AIOZ Storage's SpaceAgri case study describes a live deployment that stores IoT sensor readings, crop imagery and breeding-experiment metadata. That data feeds crop growth analysis and model training, a real dataset-and-output workload running on AIOZ Storage today. AIOZ's wider positioning is in AIOZ Storage as a DePIN backbone for Web3 data.

How AIOZ Storage Handles It

For files, you use any S3 client with the endpoint https://s3.aiozstorage.network and your access keys, and for embeddings you use the S3 Vectors client in boto3. This trimmed version of the Python vector guide creates an index and runs a similarity search:

import boto3

client = boto3.client(
    's3vectors',
    region_name='us-east-1',
    aws_access_key_id='YOUR_ACCESS_KEY',
    aws_secret_access_key='YOUR_SECRET_KEY',
    endpoint_url='AIOZ_STORAGE_ENDPOINT_URL'
)

# Create index
client.create_index(
    vectorBucketName='YOUR_BUCKET_NAME',
    indexName='my-index',
    dataType='float32',
    dimension=3,
    distanceMetric='cosine'
)

# Query similar vectors
response = client.query_vectors(
    vectorBucketName='YOUR_BUCKET_NAME',
    indexName='my-index',
    queryVector={'float32': [0.11, 0.85, 0.47]},
    topK=5,
    returnMetadata=True,
    returnDistance=True
)

The SI Memory page lists the vector API base URL as https://vector-api.aiozstorage.network with region us-east-1. To let an assistant run the same steps from prompts, connect your MCP client to https://mcp.aiozstorage.network/sse, as described on the MCP server page.

On cost, files bill at $0.017 per GB-month for storage and $0.005 per GB downloaded, while vectors bill at $0.04 per GB uploaded and $0.05 per GB-month stored. Each query costs $0.000002 plus a small charge for the data searched and returned, per the billing page.

Next step: create an access grant in the dashboard, then follow the Vector Database guide for your language.

Frequently Asked Questions

Does AIOZ Storage have a vector database?
Yes. Vector buckets and indexes work through the AWS S3 Vectors API, with Python, JavaScript and Go guides in the docs, and you store embeddings with metadata and query them by similarity.

Can I store SI training datasets on AIOZ Storage?
Yes. Datasets are objects in a bucket, so any S3 SDK, the AWS CLI or rclone can upload and read them whatever the file type is.

Does AIOZ Storage have an MCP server?
Yes. It connects over HTTP SSE at https://mcp.aiozstorage.network/sse and offers 22 tools across S3 buckets, S3 objects, vector buckets, indexes and vectors. Its upload tool returns a presigned URL rather than sending the file itself.

Can LangGraph checkpoint agent state to AIOZ Storage?
The langgraph-checkpoint-aws package accepts a custom endpoint_url in its s3_offload_config. You can set it to the AIOZ Storage endpoint, though that is LangGraph's option, not an AIOZ-built integration.

How is vector storage priced?
Uploads cost $0.04 per GB and storage $0.05 per GB-month, charged hourly. Each query costs $0.000002 plus fees for the data searched and returned, all from the same pre-paid balance.

Is AIOZ Storage connected to AIOZ AI's hosted models?
The Storage docs don't describe a connection, so treat them as separate products on the AIOZ DePIN network.

Is there a real example of AIOZ Storage used for model training data?
Yes. The SpaceAgri case study describes sensor data and crop imagery stored for crop analysis and model training.

Related Reading

• AIOZ Storage Explained: S3-Compatible Decentralized Storage - the full product model this guide builds on.
• Vector Database vs Agent Memory on AIOZ Storage - which of the two vector features fits your agent.
• AIOZ Storage Super Intelligence Memory: Record and Recall - agent memory without managing embeddings.
• What Is an MCP Server? - how the MCP standard works and what the AIOZ Storage server exposes.
• AIOZ Storage Access Control: Macaroons and Access Grants - the credentials a pipeline or agent would configure.
• File Management API with FastAPI, Boto3, and AIOZ Storage - the Python pattern most ML tooling uses for objects.
• AIOZ Storage Success Story: SpaceAgri - the deployment referenced above.

References

• AIOZ Storage Docs: Vector Database
• AIOZ Storage Docs: Vector Database, Python
• AIOZ Storage Docs: MCP Server
• AIOZ Storage Docs: Super Intelligence Memory
• AIOZ Storage Docs: Billing
• langgraph-checkpoint-aws on PyPI
• OpenAI Agents SDK: Sandbox Guide
• OpenAI Agents SDK: Sandbox Entries Reference