MCP, use Groupilot from Claude

Groupilot runs an MCP server. Connect it to your assistant and say "create my Sunday 8 pm webinar with a J-1, H-1 and replay sequence". The assistant calls the same API as everyone else, with your key or your OAuth connection, and asks before anything that reaches real people.

ConnectCopy Anchor Link

Remote server (recommended)Copy Anchor Link

In claude.ai, Claude Desktop, Cursor or VS Code, add a remote MCP server with that URL. Without a token the server answers 401 with WWW-Authenticate: Bearer realm="groupilot", resource_metadata="https://mcp.groupilot.com/.well-known/oauth-protected-resource" (on /mcp, that URL ends with /mcp): a client that supports OAuth registers itself, opens the Groupilot consent page where you pick the WhatsApp account and the rights, and gets its token. Nothing to copy. See Connect OAuth.

With a client that lets you set headers, an API key works instead: header Authorization with value Bearer gp_live_….

{ "mcpServers": { "groupilot": { "url": "https://mcp.groupilot.com/mcp" } } }

Local package (available soon)Copy Anchor Link

For clients without remote OAuth, a local bridge runs over stdio with an API key:

{
  "mcpServers": {
    "groupilot": {
      "command": "npx",
      "args": ["-y", "@groupilot/mcp"],
      "env": { "GROUPILOT_API_KEY": "gp_live_…" }
    }
  }
}

GROUPILOT_MCP_URL is optional. The package is not on npm yet: use the remote server in the meantime.

What the assistant can doCopy Anchor Link

One tool per API endpoint, with the same scopes (read, write, send), the same limits and the same errors ({ "error": { "code", "message", "field" } }). Every call appears in Settings, Integrations, Last calls, with the method MCP. There is no two-step media upload over MCP: the assistant uses import_media with a public URL.

Tools that reach real people (send_adhoc_message, create_broadcast, cancel_job, cancel_broadcast, pause_event, delete_sequence, delete_message, delete_webhook, confirm_backup_admin) are flagged destructive and their description asks the assistant to confirm with you first. Reading tools are flagged read-only.

The list above comes from the live catalogue, mcp-tools.json.

The sequence promptCopy Anchor Link

The server ships a generate_sequence prompt with arguments event_name, live_day_time, offer, and optional tone and language (French by default). It gives the assistant the rules of a sequence (a window from J-4 to J+3, at most 200 messages, H timings in whole hours, mention of all members on the two or three most important messages only) and the exact format of sync_sequence. The assistant writes, you review, then it saves with sync_sequence. Groupilot runs no AI itself: the generation happens in your assistant.

The docs as a resourceCopy Anchor Link

The server exposes groupilot://docs/llms.txt, this documentation in one file refreshed every hour, so the assistant knows the timing rules, the variables and the limits without searching. You can also paste llms-full.txt into any chat.

SafetyCopy Anchor Link

• The server has exactly the rights of the key or the OAuth grant. A key without send cannot send, whatever the assistant tries.
• Prefer read and write for everyday setup. Add send only when you want the assistant to send.
• Revoke a key from Settings, Integrations, or an OAuth connection from Connected applications: the assistant loses access within a minute.

Example conversationCopy Anchor Link

Create a recurring event "Sunday webinar", every Sunday at 20:00, communities created 3 days before. Then build a sequence with a reminder the day before at 10:00, an image one hour before, a message mentioning everyone at 19:55 with the Zoom link, and a replay message the day after at 10:00. Link it to the event and give me the smart link.

The assistant calls create_event, sync_sequence (after showing you the messages), link_sequence, then get_event and returns the smart_link.