MCP helpdesk for Claude Code, Codex and Cursor The agent that built your app can run its support.

Guardrails

It drafts. You approve.

An agent with access to your customers needs limits it cannot talk its way around. These are enforced by the server and the key, not by the prompt.

• Drafts wait for you

  The built-in Auto-Responder only proposes. Its drafted replies and articles sit in Agents Activity until you approve them, and approving is what sends.

• Every tool says what it does

  28 tools are marked read-only, so your client can let lookups run without asking. The 10 that can reach a customer are marked destructive and open-world: the signal a client uses to ask you first.

• The key sets the limits

  Give the agent a key with only the scopes it needs (tickets:read but not emails:send, say) and only the projects it should touch.

• Try it in sandbox first

  A sandbox key, or environment: "sandbox" on a single call, runs the flow on separate data. Emails are captured, never delivered.

Setup

One command in the agent you already use.

Create an API key in Helmdesk, then add the server. Full guides, including per-project config and Claude Desktop, are in the docs.

claude mcp add --env HELMDESK_API_KEY=sk_live_… \
  --transport stdio helmdesk -- npx -y @helmdesk/mcp

codex mcp add helmdesk \
  --env HELMDESK_API_KEY=sk_live_… -- npx -y @helmdesk/mcp

{
  "mcpServers": {
    "helmdesk": {
      "command": "npx",
      "args": ["-y", "@helmdesk/mcp"],
      "env": { "HELMDESK_API_KEY": "sk_live_…" }
    }
  }
}