> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zeotap.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Zeotap MCP
> Query your Zeotap CDP in plain language from Claude, ChatGPT, or any MCP-compatible AI assistant — data schema, audience and journey health, org comparisons, and suggested definitions, all read-only in this beta.
## What is Zeotap MCP?
Zeotap MCP connects your team's AI assistant directly to your Zeotap CDP — Claude, ChatGPT, or any other MCP-compatible interface — so you can ask questions about your data, audiences, journeys, and platform configuration in plain language. Instead of navigating dashboards or waiting on a data team, you ask, and the answer is grounded in your actual Zeotap data: what a field means, how a metric has moved, whether an audience or sync looks healthy, or what a good next audience might look like.
Zeotap MCP is currently available as a **Beta** feature, enabled per organisation. The **MCP server URL** used in the setup below is provided by your Zeotap team during enablement — reach out to your Zeotap contact to join the beta.
## What you can do in this beta
* **Understand your data** — schema, field-level quality, fill rate, and how key metrics and event volumes have changed over time.
* **Look up and check the health** of your audiences, journeys, destinations, sources, and identity configuration.
* **Compare configuration across your orgs** — for example, sandbox and production.
* **Brainstorm** — get suggested audience or journey definitions grounded in your existing data.
* **Turn any of the above into a shareable reference document**, on demand.
This release is **read-only** — you cannot create, edit, publish, or delete anything through MCP. See [Known limitations](#known-limitations).
## Connection setup
Pick the tab for the AI assistant your team uses.
**Admin setup** — adding Zeotap as an organization connector is done once, by a Claude **Owner** or **Primary Owner**.
In Claude, go to **Organization Settings → Connectors**, then toggle to **Organization connectors**.
Click **Add custom connector**. Enter a name (for example, `Zeotap — your org name`) and the **MCP server URL** provided by your Zeotap team during enablement.
Click **Add**. The connector is now available to everyone in your Claude organization.
**For each user**
Start a new chat, click the **+** icon, open **Connectors**, and toggle the Zeotap connector on.
On your first Zeotap-related question, sign in with your Zeotap Google account and grant the requested permissions on the consent screen.
You're connected for the session. Try one of the [sample prompts](#how-to-use-it-sample-prompts) below.
In ChatGPT, go to **Settings → Plugins**, then click **MCP**.
Click **Add MCP** and select **Streamable HTTP**. Enter a name (for example, `Zeotap — your org name`) and the **MCP server URL** provided by your Zeotap team, then click **Save**.
Click **Authenticate** next to the connector, log in with your Zeotap account in the window that opens, and grant the requested permissions on the consent screen. Once authenticated, the connector is active — try your first prompt.
Any MCP-compatible assistant can connect using the same server URL. The sign-in, consent, and permission model described on this page is identical regardless of which assistant you use.
## Querying your organizations
Your MCP connector covers your parent organization and all of its child organizations. By default, queries run against your **parent organization**; to query a specific child org instead, mention that org's name or ID directly in your prompt — for example, *"Show me active audiences for `org name`."*
## How to use it: Sample prompts
* "Where does the `loyalty_tier` field on our audiences come from — trace it back to its source."
* "Summarize our full data catalogue, calculated attributes, and audience definitions into a reference document."
* "What can I access with my current role?"
* "Document all our active audiences — their definitions, sizes, and sync destinations — as a single reference."
* "Put together a summary of our identity resolution strategy."
* "Give me the breakdown of all profile attributes with their fill rate. Why does the `X` column have a low fill rate?"
* "What event types are currently being tracked? Provide a breakdown of the total number of events for each type over the last 30 days."
* "I'm new to this org — give me a plain-language tour of what data we actually have."
* "What fields do we have on our customer profiles?"
* "Which fields are flagged as PII, and how sensitive are they?"
* "What's the fill rate for `annual_income` and `age` across our customer profiles?"
* "Show me the audiences we activated to Braze last week, and whether the campaigns actually sent."
* "What's the current status for each of our active audiences?"
* "Give me the ingestion summary for the `source name` source for the last 7 days."
* "Show me the list of all journeys in the default folder with their status summary."
* "How has the fill rate on our `purchase_amount` field moved over the last quarter?"
* "What's the trend in event volume for product-view activity over the past 30 days?"
* "List our active audiences and their current sizes."
* "What journeys are currently live?"
* "Compare the audience list in our production and non-production environments — which audiences exist in one but not the other?"
* "Are the same destinations configured in both environments?"
* "Do both environments have the same catalogue fields available?"
* "Suggest a win-back audience definition for customers who haven't purchased in 90 days."
* "What would a good cart-abandonment journey structure look like, given the data we have?"
* "Given our event data, what's a reasonable definition for an at-risk customer?"
## Known limitations
* **Read-only by design.** Creating, editing, or publishing anything in your Zeotap org is not part of this beta.
* **No change history yet.** Audience and other asset change history is planned for the next phase.
* **Historical data window.** Historical analysis of event-related data covers the last **60 days**. For anything further back, reach out to your Zeotap contact.
* **No proactive alerts.** MCP answers when asked — it does not send autonomous notifications. For automated alerting across sources and audiences, use [Alerts & Notifications](/articles/admin-customer/alerts-and-notifications) under **Admin**.
* **Rate limits.** Zeotap MCP allows up to **500 queries per minute per organization**, shared across all users querying that org, for security purposes.
* **Cross-platform correlation** with external platforms such as Braze requires an active secondary MCP connector and consistent cross-system audience identifiers — discuss your use case with your Zeotap contact.
* **Data retrieval limits.** When querying exceptionally large datasets (for example, organizations with more than 500 audiences), the assistant may return a partial set of results and invite you to request the rest in a follow-up query. This limit exists because pulling too much data at once can overwhelm your AI assistant's context window.
## Data & security
**Scope of data exposed.** MCP is built on aggregated data — schema, statistics, definitions, and configuration — not raw individual customer records. No tool in the current release returns individual-level customer identifiers or record lists, and sample values from PII-flagged fields cannot be retrieved through MCP. This is enforced at the tool level, not left to how a question is phrased.
**Data flow and processing location.** Zeotap's own MCP processing — schema, audience, and journey data handling — runs entirely on EU-based infrastructure. Separately, the LLM interface you use (Claude, ChatGPT, or others) has its own data-handling terms for the conversation itself; that is governed by the vendor's policies, and worth confirming with whichever vendor your team standardises on.
**Conversation retention.** An MCP server only ever receives a specific, structured request — a tool name and its parameters — when the assistant decides to call it. It never receives your question as asked, the model's reasoning, or the answer given back. The conversation exists entirely on the LLM side, so there is no conversation for Zeotap's MCP layer to retain; retention of prompts and responses is a question for the LLM interface in use.
**Access control.** Access is gated in two layers: an organization admin must explicitly enable the connector, and each user then authenticates individually with their own account. Every request is checked against your current Zeotap permissions, live — not just once at login. If you don't have audience-level access in Zeotap, you cannot pull audience details through MCP, regardless of how the question is phrased. If a permission check can't be completed, access is denied by default, and credentials are never included in what's returned to the assistant.
## Related Topics