Connect ChatGPT

ChatGPT MCP support is evolving across client and developer-mode surfaces. Use this page as the focused connection shape: point ChatGPT’s MCP configuration at the local icuvisor binary and keep secrets out of the model conversation.

Before you start

• Install icuvisor and run setup.
• Confirm the binary starts with icuvisor version.
• Know your non-secret athlete ID and timezone.

Stdio configuration shape

Use the same stdio server definition as the Claude clients when ChatGPT asks for a local MCP server command:

{
  "mcpServers": {
    "icuvisor": {
      "command": "/Applications/icuvisor.app/Contents/MacOS/icuvisor",
      "env": {
        "INTERVALS_ICU_ATHLETE_ID": "i12345",
        "ICUVISOR_TIMEZONE": "America/Sao_Paulo",
        "ICUVISOR_TRANSPORT": "stdio"
      }
    }
  }
}

If your ChatGPT MCP surface expects a single server object rather than a full mcpServers map, use the icuvisor object from the example as that server definition.

HTTP alternative

If your ChatGPT MCP surface expects an HTTP URL, start icuvisor with Streamable HTTP on loopback:

ICUVISOR_TRANSPORT=http /Applications/icuvisor.app/Contents/MacOS/icuvisor

Then point the client at:

http://127.0.0.1:8765/mcp

Do not bind icuvisor to a LAN address unless you intentionally want other machines to reach the unauthenticated local MCP server. The HTTP transport guide covers the security tradeoff in more detail.

Verify

Start a fresh ChatGPT conversation after saving the MCP configuration, then ask a simple profile question such as What's my FTP? The expected tool call is get_athlete_profile.