The Power My Analytics MCP server lets you connect Claude Desktop, ChatGPT, and other Model Context Protocol (MCP) clients directly to your PMA Hub so you can ask questions about your marketing data in plain language. This FAQ answers the questions we hear most often: what you can and can't do, how fresh the data is, how sign-in works, which AI clients are supported, what it costs, and how your data is handled.

The PMA MCP server is now in Beta and available to all PMA users. No early access request is required.

MCP is free to use during its Beta soft launch, with no usage-based charges, through July 31, 2026. Starting August 1, 2026, MCP usage will move to usage-based billing. The specific usage limits and pricing per plan are still being finalized; we'll share the details, and give you advance notice, before any charges begin.

Getting Started and Access

What is the PMA MCP server?

It is a hosted endpoint at https://pma-mcp.web.app that exposes your PMA Hub to Model Context Protocol–compatible AI assistants. Once connected, your AI assistant can browse your connected data sources, inspect schemas, run analytics queries against your warehoused marketing data, and read or build Data Builder datasets, all through natural-language conversation. In short, it turns your Hub into a conversational analytics surface: instead of building a report to answer a one-off question, you can simply ask.

Who is it for?

The MCP server is most useful for paid-media teams, agency analysts, and anyone running ad-hoc "what changed?", "compare X to Y," or "is anything weird?" investigations. It is designed for any PMA user with Hub access: the marketers and analysts who want answers without building a report first.

What's the difference between the PMA API and the MCP Server, and who can use each?

Both reach the same warehoused data. The MCP Server connects an AI assistant (Claude or ChatGPT) so you can ask questions in plain language; the PMA API gives your own code direct REST access. Both are now available to all PMA users. For a full breakdown, see PMA API vs. PMA MCP Server.

Are there ready-made prompts I can use?

Yes. We’ve put together pre-built prompts for every role to help you get started fast. See the PMA MCP Server Prompt Playbook.

How do I get access?

MCP is now available to all PMA users; no request or early access is needed. Simply connect your AI client using the setup guides linked at the end of this article. When you connect a browser-based AI client, your token is created automatically through the OAuth flow.

What do I need on the PMA side?

You'll need an active PMA Hub with at least one connected data source that has data syncing. That's it on the PMA side: MCP is available to all PMA users in Beta, with no special role or access request required. When you connect a browser-based AI client, your token is created automatically through the OAuth flow. For headless / server-to-server integrations, generate an API token from Hub settings (Settings > API Tokens).

Which AI clients are supported?

Any MCP-compatible client that supports the Streamable HTTP transport works. Today that includes:

• Claude Desktop, Claude Web, and Claude Code (Anthropic)
• ChatGPT (OpenAI)
• Headless / server-to-server clients (your own backend services)
• Other MCP-compatible clients that support Streamable HTTP (and, for browser-based sign-in, OAuth 2.1)

Each client has its own short setup guide, linked at the end of this article.

Do I need a paid AI plan?

For browser-based AI clients, yes: you need a Claude or ChatGPT plan tier that supports custom MCP connectors (for example, Claude Pro or Team). Individual users on a qualifying plan can add the connector themselves. In organization contexts, the AI client's organization Owner adds the custom connector once for the whole org, and individual users then connect their own PMA accounts.

In ChatGPT, you also need to select the PMA MCP app from the tools menu in each new chat, or ChatGPT will answer without using your PMA data. This is specific to ChatGPT; Claude clients invoke the connector automatically.

What You Can (and Can't) Do

What kinds of questions can I ask?

Anything that maps to the data you've centralized in your Hub. Common examples:

• Headline numbers: "How is my Facebook Ads doing this month?"
• Cross-platform comparisons: "Compare last month's spend across Facebook Ads and Google Ads."
• Period-over-period: "Is my ad spend up or down compared to last month?"
• Trends: "Show me weekly clicks for the last 12 weeks."
• Anomaly detection: "Did anything weird happen in my Google Ads spend this week?"
• Diagnostics: "Why isn't my data syncing?" or "Show me all my connected accounts and their connection status."

Behind the scenes, the server exposes 22 first-party tools grouped into discovery and diagnostics, sub-account and activity, analytics, Data Builder datasets, and template browsing. You'll rarely call them by name; your AI assistant picks the right tool based on your question. For the full inventory, see the PMA MCP Tools Reference.

Can it build or change my datasets and reports?

Yes for Data Builder datasets: the MCP server can read, write, and query your Data Builder datasets, including creating a new dataset (from a template or blank) and adding data tables to it. A dataset created through MCP is the same dataset you see in the Hub; you can edit either one in the other surface.

When the AI assistant tries to create a dataset on your behalf, you'll see a confirmation prompt asking you to approve the dataset name before anything is saved. This step is intentional, since datasets are persistent.

Configuration tasks (connecting new sources, managing sub-accounts, or changing sync schedules) are still done in the Hub. The MCP server can list and inspect those settings, but it does not change them.

Does it replace Looker Studio, Google Sheets, Excel, or the PMA API?

No. The MCP server is a new way to consume the data you've already centralized; it complements your existing tools rather than replacing them. Your Data Studio, Google Sheets, and Microsoft Excel reports stay authoritative for recurring, shareable, board-ready reporting. The PMA API and Custom-plan exports remain the path for programmatic, scheduled, and large-volume access. The MCP server proxies to the same warehoused, deduplicated, time-zone-normalized SQL layer those surfaces use; it is an additional access surface, not a replacement. It shines for ad-hoc, exploratory, single-user questions you'd otherwise build a whole report to answer.

How is this different from a single-platform AI connector?

PMA's MCP queries your warehoused, blended, multi-platform layer, so cross-platform comparison is a single tool call rather than a manual reconciliation. Because it reads the warehouse and not each platform's live API, it isn't subject to per-platform live-API rate limits or per-platform quirks.

What can't I do today?

A few things are out of scope today:

• Real-time, non-warehoused live data. The server reads PMA's warehoused layer, not platform live APIs (see Data Freshness below).
• Querying multiple hubs in one session. There is one API token per organization. If you have access to multiple hubs, disconnect and reconnect to switch between them; there is no multi-hub toggle within a single session.
• Configuration changes. Connecting sources, managing accounts, and setting sync schedules are still done in the Hub.

Data Freshness

How fresh is my data?

The MCP server queries PMA's warehoused data layer, so answers reflect the state of your warehouse after the most recent successful Data Sync for each platform. Sync cadence is plan-dependent: daily by default, with hourly refreshes available on Custom plans (with the optional hourly refresh feature). Each account's last-sync timestamp is queryable, so you (or your AI assistant) can always check current freshness by asking for your connected data sources.

How far back can I query?

The MCP layer does not currently enforce a maximum lookback window; that cap has not yet been decided. You can query historical ranges, but treat answers from very old date ranges as best-effort until the cap is formally set and documented.

Why didn't it flag the day I had zero spend as an anomaly?

This is intentional. Anomaly detection explicitly excludes zero-value days from anomaly results and reports them separately as data gaps rather than as performance drops. A zero-data day usually reflects a sync gap to backfill, not a real drop, so it's surfaced as "no data reported" instead of an anomaly.

Authentication and Security

How does sign-in work, and do I need new credentials?

No new passwords. There are two authentication paths, and both leverage credentials you already have:

• Browser-based AI clients (Claude Desktop, Claude Web, Claude Code, ChatGPT) sign in with OAuth 2.1 + PKCE against your existing PMA login (Google single sign-on or email and password).
• Headless / server-to-server clients send your existing PMA API token directly as a Bearer credential, skipping the OAuth handshake.

Both paths resolve to the same underlying PMA API token, so no separate identity is created for the MCP server.

What can the AI assistant actually access?

It can access the same warehoused marketing data your PMA login can reach in the Hub: your connected sources, schemas, analytics queries, and Data Builder datasets for the hub you authorize. It receives query results only as part of the conversation; it does not get a copy of your warehouse.

How long does my connection last?

• OAuth-issued access tokens (browser-based clients) expire 1 year after issuance. Refresh tokens are not issued, so when one expires (or if your hub access is revoked or your PMA login changes), disconnect and reconnect in your AI client to mint a new one.
• The underlying PMA API token (headless integrations) is long-lived and does not expire on its own. Rotate it from Hub settings as part of your standard production-secret rotation, then update the value in your secrets store and restart your service.

Where do I see or revoke my connections?

Open the MCP page in your Hub (left navigation). The Connected Application table lists the applications your organization has connected. (If PMA support connects to your organization to troubleshoot an issue, that connection is not shown and its usage is never counted toward your totals.) To revoke a connection, click the trash icon in the Actions column and click Confirm; that application immediately loses access, and your Hub data is unaffected. See Using the MCP Page in Your PMA Hub.

Is my marketing data sent to Anthropic or OpenAI?

Your data continues to live in PMA's existing data warehouse. No data is mirrored into the AI client; the AI client only receives the specific query results needed to answer you, as part of your conversation. Those results are then handled under your AI provider's own terms (Anthropic for Claude, OpenAI for ChatGPT). The MCP layer introduces no new PMA sub-processors beyond Google (PMA's existing hosting and warehouse provider).

Where does my data live?

Wherever it already lives in PMA: the US data center (Google Cloud, Council Bluffs, Iowa) by default, or the EU data center (Google Cloud, Frankfurt, Germany) for organizations that selected EU data residency. The MCP server does not change your data residency. For full details on how PMA handles data, see the Privacy Policy.

Pricing

Does the MCP server cost extra?

MCP is free to use during its Beta soft launch, with no usage-based charges, through July 31, 2026. Starting August 1, 2026, MCP usage will move to usage-based billing. The specific usage limits and pricing for each plan are still being finalized; we will share the full details, and give you advance notice, before any charges begin. For your current plan details, check your Hub or contact your PMA representative. Any troubleshooting activity by the PMA support team is excluded from your usage stats and will never be billed to you.

Vague prompts can increase usage on the AI client's side, because the assistant may make several discovery calls before it can answer. Specifying the platform, metric, date range, and analytic intent up front keeps those costs down. This affects your AI provider's token usage, not your PMA bill.

Limits and Known Behaviors

A few behaviors are worth knowing up front. For symptom-by-symptom fixes, see the troubleshooting article linked below.

Is there a rate limit?

Yes. 30 requests per minute per organization is a hard limit. Agentic AI clients in chatty loops (or investigations that fan out across many platforms at once) can hit it and receive a generic "too many requests" error. This is intentional, not a bug; pause for a minute, then retry, and consider tightening your prompt to reduce the number of tool calls.

Why do answers sometimes come back in different shapes?

Canonical response formats for high-frequency queries are still being standardized, so the same prompt can occasionally return slightly different structures. If consistency matters for your workflow, state the output format you want in the prompt (for example, "answer as a table with columns Platform, Spend, Clicks").

ChatGPT isn't using my PMA data. What's wrong?

ChatGPT does not invoke custom MCP connectors automatically. Open the tools menu and select PMA MCP at the start of each new chat, then resend your prompt. (Claude clients don't require this step.)

Related Articles

• Power My Analytics MCP Server Guide
• PMA MCP Tools Reference
• Connect to the PMA MCP Server from Claude Desktop
• Connect to the PMA MCP Server from Claude Code
• Connect to the PMA MCP Server from ChatGPT
• Headless/Server-to-Server Integration with PMA MCP Server
• Troubleshooting and Important Considerations for PMA MCP Server
• PMA API vs. PMA MCP Server: What Each One Does and Who Can Use It
• PMA MCP Server Prompt Playbook: Pre-Built Prompts for Every Role

Have feedback on the PMA MCP server, an idea for a new tool, or a use case you'd like us to write up? Reply to your early-access ticket or contact contactus@powermyanalytics.com with the subject line "MCP Feedback".