Everything Otto can do in the dashboard, your own AI can do from outside it. Magistry runs a remote MCP server at your dashboard domain under /api/mcp (Streamable HTTP) — connect Claude, Claude Desktop, Cursor, or any MCP client, and it gets the same tool registry the in-app assistant uses. Drive your whole store from your own AI.
Connect claude.ai
- 1In claude.ai, add a custom connector and give it the endpoint: https://app.magistry.io/api/mcp.
- 2You're routed through Magistry's OAuth flow. Pick which store the connector can see and which scopes it gets — read-only is a fine place to start.
- 3Approve, and the tools appear in Claude. Ask it something real ('what did my worst campaign spend this week?') to confirm the reads work.
Claude Desktop and other clients
Claude Desktop connects through mcp-remote. Add this to your MCP configuration:
{
"mcpServers": {
"magistry": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://app.magistry.io/api/mcp"]
}
}
}The server speaks OAuth 2.1 — authorize, token, and dynamic client registration — so any compliant client can register itself and walk the same store-and-scope flow. Clients that can't do OAuth can authenticate with an API key as a Bearer token instead.
Reads, writes, and approval cards
The full read surface is available under row-level security — the connector sees exactly what the granted store and scopes allow, nothing else. Writes come in two kinds: auto_write tools apply directly, and propose_write tools create an approval card in your dashboard — the external AI proposes, you approve in Magistry, and only then does anything execute. If you want propose_write tools to apply without the card, auto-apply is an explicit opt-in per connection in Settings → Developers. Everything is rate-limited to 120 requests per minute.
The approval card is the point: your AI can be wrong at full speed, and your store still only changes when you say so.