How to use the Stream Vessel MCP server with Claude, ChatGPT, Cursor and VS Code

Updated October 30, 2026 · 12 min read · For AWS S3, Cloudflare R2, Backblaze B2, Wasabi, DigitalOcean Spaces, MinIO, IDrive e2 and any S3-compatible storage

The Stream Vessel MCP server lets an AI assistant work with your S3 and S3-compatible storage. It can find files, read them, upload, copy, move and clean up, and it never sees your access keys. This guide covers connecting Claude, Claude Code, ChatGPT, Cursor or VS Code, choosing how much access to give, and keeping the agent safe and useful.

1. Before you start

You need three things:

1. A Stream Vessel Cloud account. Sign up free at app.streamvessel.com with Google, GitHub or an email address.
2. At least one storage connection on that account. AWS S3, Cloudflare R2, Backblaze B2, Wasabi, DigitalOcean Spaces, MinIO, IDrive e2 or any other S3-compatible endpoint. The endpoint must be reachable from the internet, because tool calls run on Stream Vessel’s servers, not on your computer.
3. An MCP client that can connect to a remote server over HTTP, such as Claude, Claude Code, ChatGPT, Cursor or VS Code.

Connections that you keep only on your device, in the Chrome extension without an account, are deliberately out of MCP’s reach: their keys never leave your browser. Connections shared with you by someone else are included.

2. OAuth or a token: which to choose

Both give the AI the same tools under the same rules. They differ in how the client proves who it is.

Either way you choose the access level, the expiry and where it works, and you can revoke it at any time.

3. Connect with OAuth

1. Add the server to your client as a remote MCP server with the URL https://mcp.streamvessel.com. You do not need a token.
   • Claude (claude.ai and the desktop app): open Settings → Connectors, choose Add custom connector and paste the URL.
   • ChatGPT: add it as a custom connector (MCP server) by its URL, where your ChatGPT plan and workspace allow custom connectors.
   • Claude Code: run the command below, then type /mcp in Claude Code, choose streamvessel and authenticate.
   • Cursor and VS Code: add a remote (HTTP) MCP server with the URL in the editor’s MCP settings; in VS Code, the MCP: Add Server command does this.
   • An agent or client of your own: add the URL the same way. The first time it signs in, Stream Vessel shows the https address it asked to send you back to; if you set the client up yourself, type its host to trust that address for your account, and the sign-in carries on. You get an email, and you can remove the address under Trusted app addresses on the MCP page, which disconnects the client. Why it works this way.
2. Sign in to Stream Vessel when your browser opens, if you are not signed in already.
3. Check the consent screen. It names the app, says where approving sends you back to, and shows the account you are approving as. Pick the access level, the expiry and where it works, then click Approve.
4. Return to your client. The tools appear there, ready to use.

Copyclaude mcp add --transport http streamvessel https://mcp.streamvessel.com

The app name on the consent screen is chosen by the app itself, and Stream Vessel has not checked it. Approve only a connection you just started yourself. If a link you did not expect opens this screen, click Deny. For a program on your own computer, such as Claude Code, the screen adds a warning that access goes to a local program; that is expected when you started the connection there.

After you approve, the app appears under Connected apps on the MCP access page, with its access level, where it works, when it was last used and when it expires. The access ends on that date; after that, connect the app again.

4. Connect with a personal token

1. Open MCP access

In Stream Vessel Cloud, open the menu and choose Account → MCP access, or go straight to app.streamvessel.com/mcp.

2. Create a token

Click New token and fill in four fields:

• Name. Say where the token will live, such as laptop-claude-code or ci-reports-readonly. You will thank yourself when it is time to revoke one.
• Scope. READ, WRITE or FULL. Start with READ; see access levels.
• Expires. 30 days, 90 days, 1 year or a custom date. Tokens always expire, after at most a year.
• Where it works. All connections, or Only these to pick a connection, a bucket and a folder.

3. Copy the token

Click Create token. The token is shown once: copy it now. Stream Vessel stores only a SHA-256 hash of it, so it cannot show the token again. If you lose it, create a new one and revoke the old one.

4. Add it to your client

Claude Code. Copy the command from the Claude Code tab and run it in a terminal. It looks like this:

Copyclaude mcp add --transport http streamvessel https://mcp.streamvessel.com \
  --header "Authorization: Bearer <YOUR_TOKEN>"

Cursor or VS Code. Click Add to Cursor or Add to VS Code next to the new token. The editor opens and offers to install the server, token included.

A project file. The .mcp.json tab gives a configuration that Claude Code reads from the root of a project. It reads the token from an environment variable, so the file is safe to commit:

Copy{
  "mcpServers": {
    "streamvessel": {
      "type": "http",
      "url": "https://mcp.streamvessel.com",
      "headers": {
        "Authorization": "Bearer ${STREAMVESSEL_MCP_TOKEN}"
      }
    }
  }
}

Copyexport STREAMVESSEL_MCP_TOKEN="svmcp_…"

Any other client. Configure a remote MCP server over HTTP with the URL https://mcp.streamvessel.com and the header Authorization: Bearer <your token>.

5. Choose the access level and where it works

Access levels

Each token or app has one level, and each level includes everything below it.

Where it works

All connections covers every connection on your account, including ones shared with you and ones you add later. Only these limits the token or app to the places you list, up to 50 of them:

• a whole connection (leave the bucket empty),
• one bucket on it (leave the folder empty), or
• one folder in that bucket, such as reports/.

Anything not listed looks to the AI as if it does not exist. A connection outside the list does not even appear in list_connections. On a connection shared with you, the AI gets the lower of its own level and the level you were given: a FULL token on a connection shared with you at READ can only read there.

6. Your first conversation

Start by asking what the AI can see. It will call list_connections, which every other tool depends on:

• “List my Stream Vessel connections and the buckets on each.”

Then try requests like these:

• Find: “Find every CSV under exports/2026/ in acme-prod-assets and summarise their headers.”
• Read: “Show me the last 50 lines of logs/app.log.”
• Measure: “How much space does backups/ take, and what is the newest file in it?”
• Edit: “Read config/app-settings.json, bump the version and write it back.”
• Organise: “Copy last month’s reports into archive/2026-09/.”
• Share: “Give me a download link for the newest backup that expires in 30 minutes.”
• Secure: “Is the acme-media-archive bucket readable by the public?”
• Save money: “Find multipart uploads older than 7 days and show them to me before cancelling them.”

7. All 27 tools

Your client lists them under the name you gave the server. Each tool is marked as read-only or as one that changes data, and clients use those marks to decide when to ask you before running it. The AI only ever sees the tools; which ones succeed depends on the access level.

8. Prompts and resources

Besides tools, the server offers three prompts: ready-made plans that the AI carries out with the tools. Clients show them in their own way; in Claude Code they are slash commands such as /mcp__streamvessel__safe_cleanup.

• audit_prefix: measures and describes what sits under a folder, changing nothing.
• summarise_bucket: describes what a bucket holds and how its storage is spread, changing nothing.
• safe_cleanup: shows you what a folder holds and asks for your confirmation before deleting any of it. Use this instead of asking for a delete directly.

Each object is also available as a resource at s3://{connection}/{bucket}/{key}, with the key percent-encoded, for clients that let you attach a file from an MCP server.

9. Tips for working safely

• Give the least access that does the job. Use READ for questions and reports, WRITE when the AI must create files, and FULL only for a clean-up you are supervising.
• Use one token per agent or machine. Then revoking one does not break the others, and Last used tells you which one is active.
• Limit it to a folder. An agent that writes reports needs reports/, not your whole account.
• Remember there is no undo. Deletes and overwrites are permanent unless your provider keeps object versions. Tools that could overwrite refuse to unless the AI passes overwrite on purpose.
• Measure before you delete. Ask the AI to run stat_prefix and show you the result first, or use the safe_cleanup prompt.
• Mind the trailing slash. logs/ is the folder; logs also matches logs-archive/. Folder tools ask for the slash.
• Ask for a search, not a crawl. “Find files named *invoice*” uses find_objects, which is much faster than browsing folder by folder.
• Keep tokens out of git. Use the environment variable in .mcp.json rather than pasting the token into the file.
• Clean up abandoned uploads. Parts of an unfinished multipart upload are billed until removed. Ask the AI to list uploads older than a few days and cancel them; uploads started in the last 24 hours are protected unless you insist.

10. Limits worth knowing

The server keeps each answer small, because an AI rereads every tool result on every later turn of a conversation. Current limits:

Each token and app is also rate-limited, so an agent stuck in a loop cannot hammer your storage. Your provider still charges its usual request and transfer fees for what the AI does.

11. Rotate, revoke and audit

• Rotate a token without downtime: create the new token, point your agent at it, then revoke the old one. A token that expires within 14 days shows a warning tag in the list.
• Revoke a token or disconnect an app with the trash icon on the MCP access page. It stops working on its next call. Your storage keys do not change, so nothing else breaks.
• Activity log. On by default, it records every tool call your tokens and connected apps make: the tool, the connection, bucket and object, and how it ended. Turning it off requires being signed in to Stream Vessel in a browser, so a stolen token cannot switch it off.

12. Troubleshooting

Still stuck? Contact us and tell us which client you use and what it reported.