Docs / Integrations

New Vintage MCP

New Vintage MCP lets an AI client securely access your winery’s data, such as orders, customers, club memberships, and campaign performance, by querying the New Vintage platform directly.

Instead of exporting a CSV and pasting it into a chat, you connect once and ask questions like:

The connected client discovers what data is available for your account, reads the schema, and writes the query. You stay in control: AI can only use tools that you permit.

Before you connect

Make sure the following are true before connecting New Vintage.

New Vintage account

An active Pro subscription. Lifetime Pro accounts qualify. Free and trial-expired accounts cannot use the MCP tools.

Connected data

At least one active integration—for example Commerce7, Klaviyo, or Mailchimp—synced into New Vintage. The MCP reads your synced New Vintage data, not the third-party APIs directly.

Sign-in

Your normal New Vintage login. Authentication happens in the browser via OAuth—you never copy or paste a token.

Connect New Vintage

New Vintage uses the Model Context Protocol with OAuth. Client controls vary; the steps below document the current Claude connection flow.

Claude setup

1

In Claude, open Settings → Connectors.

2

Find New Vintage in the directory and choose Connect.

3

A New Vintage sign-in page opens in your browser. Sign in with your usual credentials.

4

Review the access request and choose Allow access.

5

Return to Claude and enable the connector for your chat.

That’s the whole flow. There is no API key to generate, store, or rotate.

Disconnecting

You can disconnect at any time from either side.

In Claude

Settings → Connectors → New Vintage → Disconnect.

In New Vintage

Settings → Connected apps → Revoke.

ChatGPT (Coming Soon)

ChatGPT support is coming soon. This guide will be updated with connection instructions when the integration is available.

Microsoft Copilot (Coming Soon)

Microsoft Copilot support is coming soon. This guide will be updated with connection instructions when the integration is available.

Available tools

Last updated: August 14, 2026

The connector exposes four tools. All four are read-only. This list will be updated periodically as we develop more tools.

Discover Tenant Data Sources

The starting point. Reports which integrations are connected, which data sources are queryable, which tools are enabled, and how to interpret results.

List Tenant SQL Sources

Lists the approved data objects currently queryable for your account.

Describe Tenant SQL Sources

Returns schema details—columns, tenant key, time columns, and join relationships—so the connected assistant can write a correct query without probing your tables.

Safe Tenant SQL

Runs a single validated, read-only, tenant-scoped SELECT against approved data sources and returns the rows.

Security and data handling

This section is written for IT and security reviewers.

Connection

Endpoint

https://mcp.newvintage.ai/mcp — HTTPS only.

Protocol

Model Context Protocol over streamable HTTP.

Browser origins

The server accepts cross-origin requests only from https://claude.ai and https://claude.com.

Authentication

Authentication is OAuth 2.0 with dynamic client registration, brokered through WorkOS AuthKit—your organization’s existing identity provider and SSO configuration apply. No static, shared, or long-lived API keys are used.

Access tokens expire after 60 minutes.

Refresh tokens are issued per user, tracked as a family, and rotated on use. Rotation is enforced; a replayed refresh token revokes the whole family.

The token endpoint, authorization endpoint, and registration endpoint are all rate-limited per IP and globally.

RFC 7009 token revocation is supported at /oauth/revoke.

Tokens are audience-bound: a token minted for one environment will not validate against another.

Authorization

Every tool call re-checks, on each request:

1

The bearer token is valid and its audience and algorithm match.

2

The session holds the mcp:read scope. A session with no scope record is treated as read-only—the code never infers write or admin access.

3

The caller has active membership in the tenant being queried. Tenant identity is derived from the OAuth session, via WorkOS organization or user membership, not from a value the model supplies. A tenantId argument can only disambiguate among tenants you already have access to; it cannot grant access to one you don’t.

4

The tenant holds a paid Pro entitlement.

Query safety

Safe Tenant SQL is the only tool that runs a query, and it is constrained on several independent layers:

Statement type

Only a single SELECT or WITH statement is accepted. Mutations, DDL, multiple statements, and unsafe functions are rejected before execution.

Object allowlist

The query is parsed and every referenced object is checked against the approved data-source catalog for your tenant. Sensitive and internal objects are excluded.

Mandatory tenant scoping

The query must include a quoted ‘__TENANT_ID__’ placeholder, which the server substitutes with the tenant resolved from your session. The model never sees or supplies a raw tenant identifier.

Validated equals executed

The statement that is executed is the same statement that passed validation; it is audited to prevent a validate-then-swap gap.

Database-level isolation

Execution is downgraded into a least-privilege Postgres role, agent_sql_executor, with NOLOGIN and NOBYPASSRLS, so row-level security policies—not just application-layer parsing—bound the result to your tenant. Even a validator bypass would not return another tenant’s rows.

Result sampling

Large result sets are returned as a prompt-safe preview rather than in full. When this happens, the response is explicitly flagged with resultMode, resultWindow, and a notice telling the connected client the rows are examples of shape and content only, and that totals, rankings, percentages, and exports must come from a narrower or aggregate query. This prevents the assistant from reporting a confident number derived from a truncated sample.

What is logged

New Vintage records product analytics on MCP usage—which tool ran, whether it succeeded, the category of any failure, which data objects were referenced, row counts, and result mode.

Analytics events are sanitized before capture: email addresses are redacted, raw SQL statements are stripped, and the free-text intent field each tool call carries is scrubbed and flagged if redaction occurred. Query result values are not written to the analytics stream.

Data flow

The MCP reads from New Vintage’s own database—the data you have already synced into the platform. It does not call Commerce7, Klaviyo, or Mailchimp APIs during a query, and it does not write to them. Query results flow to the connected client to answer your question. For Claude, they are subject to Anthropic’s data handling terms for your Claude plan.

Anthropic commercial terms →

Limits and known behavior

Pro only

Free-tier accounts receive a clear ‘requires a Pro account’ error rather than partial results.

Synced data only

The connector sees what your last integration sync brought in. Discovery reports each integration’s last sync time and status so you can tell whether data is current.

Read-only by design

The connected assistant cannot place an order, edit a customer, send a campaign, or change any setting through this connector.

One statement per call

Complex analysis is performed as several sequential queries, not one batched script.

Previews are not exports

For a complete data export, use New Vintage’s export features rather than asking the assistant to page through results.

Troubleshooting

Requires a Pro account

The tenant’s subscription is not active Pro. Check billing in New Vintage settings.

Sign-in loop, or the connector never prompts for OAuth

Remove the connector and re-add it. The authorization flow starts from the add or reconnect action; it will not begin from within an existing chat.

Access denied: this token lacks the mcp:read scope

The session predates a scope change. Disconnect and reconnect to re-authorize.

The client reports a number that looks wrong

Check whether the result was a sampled preview. Ask the assistant to re-run the question as an aggregate query—COUNT, SUM, or GROUP BY—rather than counting returned rows.

No data sources listed

No integration has completed a sync yet. Connect and sync an integration in New Vintage first.

Support and policies

Find assistance, technical references, and policy information below.

AI intelligence for wineries
of all sizes.

COMPANY

About us

RESOURCES

Help Center (Coming soon)

© 2026 New Vintage Labs · Napa, California

AI intelligence for wineries
of all sizes.

COMPANY

About us

RESOURCES

Help Center (Coming soon)

© 2026 New Vintage Labs · Napa, California

AI intelligence for wineries
of all sizes.

COMPANY

About us

RESOURCES

Help Center (Coming soon)

© 2026 New Vintage Labs · Napa, California