Skip to main content
Beyond the REST API, Vellaro exposes two Model Context Protocol (MCP) servers on mcp.vellaro.io. An AI agent can query and operate the store with the same auth, permissions, and tenant isolation as the HTTP endpoints documented here.

Endpoints

Transport is streamable-HTTP (stateless). No secret is stored: the credential travels in the MCP request headers and is forwarded verbatim to the REST API.

Storefront MCP

Authenticated with the store’s publishable key (cx_pk_live_...) — it identifies the store and cannot touch the /admin/* endpoints.

Admin MCP

Authenticated with a staff token + granular permissions; every action is scoped to its store. Mutations require confirm=true and are audit-logged.

Connecting

Header-based clients (Claude Desktop, Claude Code)

The recommended way: the server endpoint plus the credential in the headers.

claude.ai connector (key in the URL — storefront only)

The claude.ai “custom connector” dialog supports OAuth only, not custom headers. For the storefront you can pass the publishable key directly in the URL:
Leave the OAuth fields (Client ID / Secret) empty. The header still takes precedence: if the client sends X-Vellaro-Key, the query string is ignored.
Storefront only. The publishable key is read-only (like a Stripe pk_) and is acceptable in a URL, but it will appear in the server access logs. Admin stays header-only: a staff token must never travel in a URL.
The MCP servers are a thin layer over this same REST API: anything a tool does, you can do over HTTP. The endpoints listed in this section remain the source of truth. See also Authentication.