CDP MCP

The CDP MCP server is a dedicated Model Context Protocol server for your LiveRecover Customer Data Platform. It lets an AI client like Claude explore your customer data — profiles, journeys, audiences, and behavioral insights — all scoped to the single brand you connect. The server registers 32 CDP tools in total; this page documents the 16 core read tools that are available on the standard CDP entitlement, with no beta flag required. The additional families (commerce, insights, extended lookup/query, and the write / ingest / privacy tools) are covered under Beyond the core read tools.

It runs on its own endpoint (https://cdp-mcp.vyg.app/mcp), separate from the conversation-focused LiveRecover MCP. Connect to it when you want to ask questions about who your customers are and how they behave, rather than about individual support conversations.

What you can do

• Look up a customer — pull a single customer's consolidated profile (identities, spend, tags) or their full chronological journey across orders, messages, and browse activity.
• Search and filter your base — find customers by email, phone, or name, or filter your whole base by tags, spend, order count, recency, and engagement to answer questions like "top 50 spenders" or "high-value customers who have gone quiet."
• Read raw events and sessions — inspect a profile's underlying event history and web sessions.
• Work with audiences — list and read your saved segments, and preview how many customers match an ad-hoc audience condition before anything is saved.
• Analyze behavior — get a one-call snapshot of your base, trend a single event type over time, compute a browse-to-purchase funnel, or run sum/avg/min/max metrics over a property.
• Discover your schema — list the event types and properties available for your brand so your AI client can build valid queries without guessing.

The tools documented on this page are read-only and scoped to one brand. The server also registers beta-gated write, event-ingestion, and privacy tools — off by default and fail-closed behind a separate write entitlement — which are described under Beyond the core read tools rather than here.

Core read tools

Beyond the core read tools

The server registers 32 CDP tools in total. The 16 above are the core read tools available on the standard CDP entitlement. The remaining 16 are intentionally not documented on this page — they are either gated behind beta entitlements or documented separately, so their absence here is deliberate, not an omission:

• Commerce (cdp_commerce_orders, cdp_commerce_subscriptions, cdp_commerce_products) — brand-scoped reads over synced Shopify orders and subscription contracts plus the live products proxy. Requires the commerce entitlement and a connected Shopify integration.
• Insights (cdp_insights_ltv, cdp_insights_rfm, cdp_insights_products, cdp_insights_at_risk, cdp_customer_insights) — brand-wide LTV / RFM / product / at-risk aggregates and a per-customer insight read. Requires the commerce entitlement.
• Extended lookup & query (cdp_customer_lookup, cdp_list_integrations, cdp_query_aggregate, cdp_list_profile_aliases) — the unified customer record by any identifier, connected-integration listing, bucketed aggregates over the scoped query path, and profile-alias reads. (cdp_customer_lookup and cdp_list_integrations also sit behind commerce entitlement.)
• Write, ingest & privacy (cdp_profile_write, cdp_segment_write, cdp_events_ingest, cdp_privacy) — the only tools that mutate data. They are fail-closed behind the write entitlement (the GDPR delete/anonymize privacy ops additionally require staff access or privacy entitlement), so they are inert unless a brand is explicitly granted the flag.

If none of these families is enabled for your brand, the server behaves exactly as the read-only surface documented above.

New here? Start with the Customer insights quickstart.

Connect

The CDP MCP server speaks HTTP transport at:

https://cdp-mcp.vyg.app/mcp

It is POST-only (no SSE / streaming transport). A development endpoint is available at https://cdp-mcp.dev.vyg.app/mcp.

You can authenticate two ways — an interactive OAuth 2.1 login, or a vyg_ API key. Both resolve to the same brand-scoped access (the core tools are read-only; any write tools stay fail-closed behind the write entitlement regardless of how you authenticate).

Option 1 — OAuth 2.1 (interactive clients)

This is the same login flow as the main LiveRecover MCP. When your client first calls a CDP tool, a browser opens to the LiveRecover login page; you log in, select a brand, and approve. Your session is then scoped to that brand and tokens refresh automatically. See Authentication for the full flow.

Claude Code:

claude mcp add cdp-mcp --transport http https://cdp-mcp.vyg.app/mcp

Claude Desktop / Cursor — add the server to your MCP config:

{
	"mcpServers": {
		"cdp-mcp": {
			"type": "http",
			"url": "https://cdp-mcp.vyg.app/mcp"
		}
	}
}

Option 2 — vyg_ API key (headless / programmatic)

If you already have a per-brand vyg_ API key — the same key class used by the CDP REST API — you can pass it as a bearer token instead of running the OAuth flow. This suits headless clients, scripts, and shared automations.

A key is issued from your web-client settings (self-serve), or by a Voyage staff member on request. It is minted off your brand's connected Shopify shop, and is shown exactly once at creation — store it in a secret manager.

Pass the key in the Authorization header:

Authorization: Bearer vyg_your_key_here

Claude Code:

claude mcp add cdp-mcp --transport http https://cdp-mcp.vyg.app/mcp \
  --header "Authorization: Bearer vyg_your_key_here"

A vyg_ key already carries its brand scope, so there is no brand-selection step — every tool call operates on that brand.

Access requirement

The CDP tools are gated by your brand's CDP entitlement. If CDP is not enabled for your brand, every tool returns a clear CDP is not enabled for this brand. error and no data. Contact your LiveRecover account team to enable it.

Good to know

• Read-only by default. The tools on this page never write, save, or delete anything, and audience previews are counts only — nothing is persisted. The server's separate write / ingest / privacy tools are fail-closed behind the write entitlement (see Beyond the core read tools).
• One brand per connection. All access is scoped to the brand resolved from your credential. To work with a different brand, connect with that brand's credential.
• Your brand only. Every query is bound to your brand's scope server-side. A tool can never return another brand's customers, events, or segments.