What Is an MCP Server? How AI Agents Access Storage
An MCP server is a program that exposes one system's tools and data through the Model Context Protocol, so any MCP-compatible assistant can use them without a custom connector. For storage, that means you can ask an assistant to list a bucket, read a file or run a vector search.
TL;DR:
- MCP is an open-source standard for connecting super intelligence (SI) apps such as Claude or ChatGPT to outside tools and data through one protocol.
- MinIO's AIStor MCP Server (in tech preview) and Microsoft's Azure MCP Server already let agents manage object storage through conversation.
- AIOZ Storage ships an official MCP server at
https://mcp.aiozstorage.network/ssewith 22 tools across S3 buckets, S3 objects, vector buckets, indexes and vectors.
Updated Oct 19, 2026: corrected the AIOZ Storage SSE URL to https://mcp.aiozstorage.network/sse and added the S3 bucket and object tools that the MCP server docs now list.
Updated Oct 22, 2026: merged our separate post How to Build an MCP Server for AIOZ Storage, Step by Step into this guide, since the official server now covers files. We kept only when a custom server still makes sense.
What Is an MCP Server?
The official MCP documentation calls MCP "an open-source standard for connecting AI applications to external systems" and compares it to "a USB-C port" for those apps. An MCP server is the piece that exposes one specific system, such as a database, a file store or an API, through that standard interface. Instead of every assistant needing its own connector for every tool, both sides speak MCP once and you plug them together.
Microsoft's Azure documentation names three roles you will see everywhere. The host is the SI application you talk to, and the client is the part inside the host that manages the connection. The server is the program that exposes tools, data and prompts.
Why Does Storage Need an MCP Server?
Storage is one of the first things an agent needs to touch, because it has to read input files, save output and organize results. Before storage MCP servers existed, you gave an agent that access through custom scripts or a separate integration for each platform. The MCP docs say the standard "reduces development time and complexity when building, or integrating with" an agent, and storage is where you feel that first.
A Real Example: MinIO's AIStor MCP Server
MinIO, an established S3-compatible platform, offers an AIStor MCP Server that it still labels as a tech preview. According to MinIO, an agent can "create buckets, upload and download objects, add tags and metadata, or convert text directly to stored objects." It can also create users, groups and policies, manage lifecycle rules and check cluster health.
MinIO pitches it as a way to run admin tasks "without memorizing CLI commands or writing scripts." If you run your own AIStor cluster, that is the problem it removes for you.
Another Real Example: Azure MCP Server
Microsoft's Azure MCP Server applies the same idea to a whole cloud, with Azure Storage as one of many resources. Microsoft says it "enables AI agents and clients to interact with Azure resources using natural language commands." You can connect to it from Visual Studio Code, Visual Studio, Cursor, Windsurf, IntelliJ and other editors, so it lives where developers already work.
Does AIOZ Storage Have an MCP Server?
Yes. You add it to an MCP-compatible client over HTTP SSE at https://mcp.aiozstorage.network/sse, and the client then shows the AIOZ Storage S3 and vector tools. There is no login step, because every tool call carries an access key ID and secret access key that AIOZ Storage checks before it runs the call. AIOZ Stream ships its own server too; see what the AIOZ Stream MCP server exposes.
The MCP server page lists 22 tools in five groups:
- S3 buckets (3):
list_s3_buckets,create_s3_bucket,delete_s3_bucket - S3 objects (6):
list_s3_objects,get_s3_object,head_s3_object,put_s3_object,copy_s3_object,delete_s3_object - Vector buckets (4):
create_vector_bucket,list_vector_buckets,get_vector_bucket,delete_vector_bucket - Vector indexes (4):
create_index,list_indexes,get_index,delete_index - Vectors (5):
put_vectors,get_vectors,list_vectors,delete_vectors,query_vectors
Two details change how you use the object tools. put_s3_object returns a presigned upload URL that lasts 60 to 3,600 seconds (900 by default), so your client still sends the file itself. get_s3_object reads up to 4 GiB of content per call, the same limit as the S3 gateway.
On the vector side, an index can have up to 4,096 dimensions and one put_vectors call takes up to 500 vectors. Each vector carries up to 40 KB of metadata across 50 keys. query_vectors returns 1 to 100 results (10 by default) using cosine, Euclidean or dot-product distance. These are the same vector buckets as the AIOZ Storage Vector Database, exposed as tools.
Because the official server now covers files as well as vectors, you rarely need to write your own. A custom server, built with the official MCP SDKs on top of an S3 client, still makes sense in two cases. Your workflow needs a tool outside these 22, or you want keys held by a server you run instead of passed in each tool call.
What the Docs Don't Cover Yet
The docs don't publish rate limits specific to MCP calls, and they don't say which MCP clients AIOZ has tested against the server.
How AIOZ Storage Handles It
You need two kinds of credentials, depending on the tools you call. S3 tools use credentials from an S3 Access Grant, and a grant scoped to one bucket is enough for object work in that bucket. Vector bucket tools need a Vector Access Grant with All Buckets scope. The docs advise credentials with only the permissions the task needs. An assistant that only reads files should get a grant for one bucket without Write or Delete. Once your client is connected, a request to list objects under a prefix sends these arguments to list_s3_objects:
{
"S3_ACCESS_KEY_ID": "your-access-key-id",
"S3_SECRET_ACCESS_KEY": "your-secret-access-key",
"bucketName": "your-bucket",
"prefix": "docs/",
"maxKeys": 100
}If the response has isTruncated set to true, you pass nextContinuationToken back as continuationToken to get the next page. In practice you rarely type this JSON yourself. You ask "List the objects under docs/ in your-bucket, then read docs/overview.txt" and the assistant fills in the calls.
Next step: create an access grant in the AIOZ Storage dashboard, then add https://mcp.aiozstorage.network/sse to your client as described on the MCP server page.
Frequently Asked Questions
What is an MCP server in simple terms?
It is a program that exposes one system's tools and data, such as a database, storage platform or API, through the Model Context Protocol. Any MCP-compatible assistant can then use it without a custom integration.
Why do storage platforms build MCP servers?
They let your agents manage storage through conversation instead of CLI commands, console screens or custom scripts. MinIO names exactly that problem as the reason for its AIStor MCP Server.
Does AIOZ Storage have an MCP server?
Yes. It connects over HTTP SSE at https://mcp.aiozstorage.network/sse and exposes 22 tools for S3 buckets, S3 objects, vector buckets, indexes and vectors, with credentials passed on each call.
Can the AIOZ Storage MCP server upload files?
It creates the upload link. put_s3_object returns a presigned PUT URL with its headers and expiry, and your client or script then uploads the file to that URL.
Which storage platforms have MCP servers?
MinIO has its AIStor MCP Server in tech preview, Microsoft covers Azure Storage in its Azure MCP Server, and AIOZ Storage has its own server for S3 and vector work.
Who created the Model Context Protocol?
MCP is an open-source standard. Its docs show it working across SI clients such as Claude and ChatGPT and developer tools such as VS Code and Cursor.
What is the difference between an MCP host, client and server?
In Microsoft's terms, the host is the SI application, the client is the part of the host that manages the connection, and the server exposes tools and data such as Azure Storage or AIOZ Storage.
Related Reading
- Vector Database vs Agent Memory on AIOZ Storage - the vector buckets the MCP vector tools work on, and how they differ from Memory.
- Object Storage Vector Database: Where AIOZ Storage Fits - what the vector buckets behind the MCP vector tools are, and when they fit.
- AIOZ Storage for SI and Data Pipelines - the hub for datasets, checkpoints and vectors on AIOZ Storage.
- AIOZ Storage Super Intelligence Memory: Record and Recall - a direct memory API, a different path than calling AIOZ Storage as an MCP tool.
- AIOZ Storage Developer Platform: SDKs, CLI, and Migrations - the SDK and CLI route for scripted file work.
- What Agent Sandboxes on Modal and E2B Store - sandbox mounts, another way agents reach object storage.
- Offloading LangGraph Agent Checkpoints to AIOZ Storage - an agent-framework integration that does not use MCP at all.
- AIOZ Storage llms.txt: Docs Built for Super Intelligence - the machine-readable docs files AIOZ publishes for coding assistants.
References
Related Content
What Is Distributed Storage? How It Works, With Examples
Distributed storage keeps data on many independent nodes that act as one system. See how it works, replication vs erasure coding, its types and trade-offs.
What Are Egress Fees? How Download Charges Work in Storage
Egress fees are charges for data leaving a storage provider's network. See what triggers them, how to estimate them, and what AIOZ Storage's docs say.
Object Storage Vector Database: Where AIOZ Storage Fits
An object storage vector database keeps vectors in the storage layer, not a cluster. See how it differs, when to pick it, and where AIOZ Storage fits in.
AIOZ Storage Vector Database Pricing: What You Pay For
AIOZ Storage vector database pricing line by line: upload, storage, queries, data searched and data returned, with a worked monthly example you can reuse.
How to Create a Bucket and Upload a File on AIOZ Storage
Create a bucket and upload a file on AIOZ Storage in five steps: sign up, set a passphrase, add credits, and upload. Plus what the passphrase cannot recover.
AIOZ Storage Rate Limit: Working Within 100 Requests/s
The AIOZ Storage rate limit is 100 requests per second. See what the docs do and don't say, and how to stay under it with boto3 retries and a limiter.