Skip to content

AI connections

AI connections link an AI assistant to your Spheron account through the Model Context Protocol (MCP), the open standard assistants use to call external tools. Once connected, you deploy, check, and manage GPU instances by asking in chat. This page covers how to connect each supported client, what an assistant can and cannot do, and the safeguards on every connection.

There is no API key to create or paste. You sign in with your Spheron account once, and the connection acts as you. A connection added in Claude works in Claude on the web, the desktop app, and the mobile app.

Prerequisites

  • A Spheron account with credit on it. See Billing.
  • At least one SSH key saved in Settings > SSH Keys, so the assistant can launch instances you can log in to. See User settings.
  • An assistant that supports remote MCP servers with OAuth sign-in.

Connect an assistant

Every client uses the same connector URL:

https://connect.spheron.ai/mcp

The URL, a Copy button, and the steps below also appear in Settings > AI connections at app.spheron.ai.

Claude

  1. In Claude, open Settings > Connectors > Add custom connector.
  2. Paste the connector URL. Leave the client ID and secret empty.
  3. Sign in to Spheron and choose a spend limit.

ChatGPT

  1. In ChatGPT, open Settings > Apps and connectors > Advanced and turn on developer mode.
  2. Click Create, paste the connector URL, and choose OAuth for authentication.
  3. Sign in to Spheron and choose a spend limit.

Codex CLI

  1. Run these two commands:

    codex mcp add spheron --url https://connect.spheron.ai/mcp
codex mcp login spheron

  2. Sign in to Spheron in the browser that opens and choose a spend limit.

Hermes Agent

  1. Add this entry to ~/.hermes/config.yaml:

    mcp_servers:
  spheron:
    url: "https://connect.spheron.ai/mcp"
    auth: oauth

  2. Run hermes mcp login spheron, sign in to Spheron in the browser that opens, and choose a spend limit.

  3. Start a new Hermes session, or run /reload-mcp.

Cursor

  1. In Cursor, open Settings > MCP > Add new MCP server.
  2. Choose HTTP (streamable) and paste the connector URL.
  3. Click Connect, sign in to Spheron, and choose a spend limit.

VS Code

  1. Open the Command Palette, run MCP: Add Server, and choose HTTP.
  2. Paste the connector URL.
  3. Sign in to Spheron when prompted and choose a spend limit.

Other MCP clients

Add the connector URL in any client that supports remote MCP servers with OAuth. Sign-in must finish on your own computer or in a supported app (Claude, ChatGPT, VS Code, or Cursor).

What the assistant can do

  • Find capacity. Search GPU offers and CPU nodes by model, count, price, and region.
  • Manage instances. Launch, stop, start, restart, rename, and terminate GPU instances and CPU nodes.
  • Diagnose problems. Ask why an instance is stuck or failed, and the assistant checks its current status and explains it.
  • Manage storage and access. Create, resize, attach, detach, and delete volumes, and add or delete SSH keys.
  • Work with teams. Create a team, and accept or decline invitations.
  • Read your account. Check balance, usage, and transactions.

Example requests:

Find the cheapest H100 with at least 2 GPUs in Europe and deploy it with my laptop SSH key.
Why is my instance "train-run-3" still deploying?
Stop every instance I am not using and tell me what it saves per hour.

What the assistant cannot do

A connection acts as you, on your account and the teams you belong to. It cannot:

  • Add funds or change your payment card.
  • See or create API keys.
  • Invite or remove team members.
  • Raise its own spend limit.
  • Reach any admin function.

Safeguards

Spend limit per connection

When you connect, choose how much the instances launched through that connection may cost per hour, all running at once: $5, $20 (preselected), $50, $100, a custom amount from $0.01 to $100,000, or No limit (your balance is the only cap).

A launch or start that would take the connection over its limit is refused. Volume storage is not counted in the limit. Only you can change the limit, from Settings > AI connections. The assistant cannot raise it.

A preview before anything that spends or deletes

These actions return a preview instead of acting:

  • Launching, starting, stopping, restarting, or terminating an instance
  • Creating, resizing, attaching, or deleting a volume
  • Saving an SSH key

For a launch, the preview shows the price, the SSH key the instance uses, and the spend limit it counts against. The assistant is instructed to show you the preview and wait for your yes before confirming. A confirmation sent within seconds of its preview is refused, so an assistant cannot approve its own action in the same breath.

Each connection can carry out at most 10 confirmed actions per hour.

You see who is asking

The Spheron sign-in page names the app that is connecting and where its sign-in finishes: on your computer, or in Claude, ChatGPT, VS Code, or Cursor. Check both before you click Allow. An app with a misleading name stands out because its sign-in destination does not match.

Manage connections

Open Settings > AI connections to see every connection you approved. Each entry shows:

  • The client name and where its sign-in finished
  • When it was connected and when it was last used
  • Its spend limit, and what instances launched through it cost per hour right now

Change a limit from the dropdown on the entry. Click Disconnect to remove one connection, or Disconnect all to remove every connection. The assistant loses access immediately. Instances it launched keep running, so stop or terminate them from Instances if you no longer need them.

What's next

Troubleshooting

Issue: A launch is refused for the spend limit

Symptoms: The assistant reports that the deployment would exceed the connection's spend limit.

Diagnosis: The instances already running through this connection, plus the new one, cost more per hour than the limit.

Resolution: Stop or terminate instances launched through the connection, or raise the limit in Settings > AI connections.

Issue: A deploy fails because of the SSH key

Symptoms: The deploy is refused with a message about a missing or unusable SSH key.

Diagnosis: Your account has no SSH key, or the saved key has no valid public key data.

Resolution: Add a valid public key in Settings > SSH Keys, or ask the assistant to save one, then deploy again. No failed instance is left behind.

Issue: The client cannot finish sign-in

Symptoms: The browser shows an error after you sign in, or the client never receives the connection.

Diagnosis: The client finishes sign-in somewhere other than your own computer or a supported app.

Resolution: Use one of the clients listed on this page, or a client that completes OAuth on localhost.

Powered by Spheron
© 2026 Spheron Network. All rights reserved.