Connect an AI agent

The Centaur MCP server gives your AI client read-only Centaur tools. You sign in with your Centaur account, not an API key.

Add the server

Every client uses this server URL:

Server URL

https://mcp.centaur.io/mcp

Claude
ChatGPT
Cursor
Codex
Claude Code

Go to Customize > Connectors, select Add, then select Add custom connector.
Enter the name Centaur and the server URL, then select Continue.
Keep the detected settings, select Continue, then select Add.
Sign in to Centaur and select Allow access. If sign-in does not open, select Connect on the connector.
In a chat, open + > Connectors and turn on Centaur.

On a Team or Enterprise plan, an owner adds the connector under Organization settings > Connectors. Each member then selects Connect on it under Customize > Connectors. Anthropic’s custom connector guide has more detail.

Your account and workspace policies decide whether you can add custom MCP servers.

Open ChatGPT Plugins, select Add, then select Add custom MCP server.
Enter the name Centaur. Under Connection, choose Server URL and enter the server URL. Keep OAuth as the authentication.
Wait a few seconds for discovery, then open Advanced OAuth settings. Keep Client Identifier Metadata Document (CIMD) as the registration method and keep the ticked default scopes.
Tick I understand and want to continue, then select Create as a plugin.
Select Connect Centaur, sign in to Centaur, and select Allow access.

If the plugin shows no tools, go to Settings > Plugins > Centaur and select Refresh tools. OpenAI’s custom MCP server guide has more detail.

These steps are for the Cursor desktop app. Cursor needs a public client ID, which Centaur gives you for the environment you connect to. It identifies Cursor, not you, so you still sign in. You need no client secret.

Add this entry to ~/.cursor/mcp.json. Keep your other servers, and replace <CENTAUR_CURSOR_CLIENT_ID>.

~/.cursor/mcp.json

{
  "mcpServers": {
    "centaur": {
      "url": "https://mcp.centaur.io/mcp",
      "auth": {
        "CLIENT_ID": "<CENTAUR_CURSOR_CLIENT_ID>",
        "scopes": [
          "events.read",
          "messages.read",
          "feed.read",
          "positions.read",
          "directory.read",
          "stats.read",
          "summaries.read",
          "offline_access"
        ]
      }
    }
  }
}

Restart Cursor.
Open Customize in the sidebar and connect Centaur.
Sign in to Centaur and approve access.

Add the server:

Terminal

codex mcp add centaur --url https://mcp.centaur.io/mcp

Sign in to Centaur and approve access:
Terminal
codex mcp login centaur
Run codex mcp list and check that centaur is listed.

To keep the server in your config file instead of step 1, add this entry:

~/.codex/config.toml

[mcp_servers.centaur]
url = "https://mcp.centaur.io/mcp"

OpenAI’s Codex MCP guide has more detail.

Add the server:

Terminal

claude mcp add --transport http centaur https://mcp.centaur.io/mcp

Sign in to Centaur and approve access:
Terminal
claude mcp login centaur
You can also sign in from a Claude Code session with /mcp.
Run claude mcp list and check that centaur shows as connected.

Anthropic’s Claude Code MCP guide has more detail.

Check the connection

Ask your client:

Prompt

Use Centaur to call list_events with limit 5 and summarize the response.

The client should show a successful list_events call, even if it returns no events. That proves the server, your sign-in, and events.read work. It does not test every scope or token refresh. Example prompts has more to try.

Add the Centaur skill

The optional Centaur skill gives agents a guided workflow for Centaur’s tools. It uses the connection you already added, so it does not sign in or create keys.

Terminal

npx skills add https://github.com/centaur-io/centaur-partners-skill

Troubleshoot

Problem	Fix
Centaur does not appear in the client	Restart the client, or open a new chat after you turn on the connector. In Codex and Claude Code, check `mcp list` and run the add command again.
Sign-in does not open, does not finish, or has expired	Repeat the sign-in step for your client. In Claude, ChatGPT, or Cursor, disconnect and connect again. If that fails, remove the server and add it again with the plain server URL. In Cursor, keep the client ID in the entry.
Cursor sign-in fails	Check that the client ID is the one for the environment you connect to.
ChatGPT asks for a client ID or secret	Wait for discovery to finish, then check that Registration method under Advanced OAuth settings is CIMD and the server URL ends in `/mcp`.
A tool is missing or returns a permission error	A connection keeps the scopes it had when you approved it. Disconnect and connect again to pick up newer scopes.
A tool reports a used-up quota	All your clients share one hourly MCP quota. See MCP quota.

Build your own client

Each tool mirrors a REST endpoint: Data reads maps them, and the endpoint’s reference page documents the arguments and results. Centaur MCP uses standard MCP OAuth discovery:

Item	Value
Resource	`https://mcp.centaur.io/mcp`
Protected-resource metadata	`https://mcp.centaur.io/.well-known/oauth-protected-resource/mcp`
Authorization server and scopes	Listed in the protected-resource metadata
Client registration	Client ID Metadata Document (CIMD), Dynamic Client Registration, or a pre-registered public client ID
Refresh tokens	Register the `refresh_token` grant and request `offline_access`
Token audience	The resource URL
Token transport	`Authorization: Bearer` header

Prefer CIMD: it needs no registration call. MCP failures covers token and scope errors. Some non-interactive integrations, such as OpenAI API-side MCP calls, cannot run the sign-in flow. Your application must obtain the token and supply it.

Get started

Embed widgets

Build with the API

Receive webhooks

Connect AI agents

Add the server

Check the connection

Add the Centaur skill

Troubleshoot

Build your own client