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 guide walks you through what the MCP server is, how it fits with the rest of PMA, how to connect, what tools are available, and how to put them to work, with worked examples covering both basic and advanced use.

The PMA MCP server is currently in Alpha. Access is granted per organization on request. If you would like to test it, contact your PMA representative or open a ticket with the subject line "MCP Early Access" at support.powermyanalytics.com.

What Is the Power My Analytics MCP Server?

The Power My Analytics MCP server 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.

You do not need new credentials. AI assistant clients authenticate using OAuth 2.1 with PKCE against your existing PMA login (Google single sign-on or email and password); headless / server-to-server integrations authenticate with your existing PMA API token as a Bearer credential. Either way, your data stays in PMA's existing data warehouse; the AI assistant or your service only receives the specific query results it needs to answer you.

In short: the MCP server turns your PMA Hub into a conversational analytics surface. Instead of building a Looker Studio scorecard or assembling a Data Builder dataset to answer a one-off question, you can simply ask.

How the MCP Server Fits with the Rest of Power My Analytics

The MCP server is a new way to consume the data you have already centralized in your PMA Hub. It does not replace your existing reporting tools; it complements them. Here is how it lines up with the rest of the platform:

The MCP server shines when you want to ask an ad-hoc question ("How is my Facebook Ads performance compared to last month?", "Did anything weird happen this week?", "Which campaigns are driving the most conversions?") without first having to build a report.

Requirements, Prerequisites, and Limitations

Requirements on the Power My Analytics Side

• An active PMA Hub with at least one connected data source that has data syncing.
• Org Admin role in the PMA Hub.
• Alpha access granted by Power My Analytics. Alpha access is allow-listed per organization. To request access, contact your PMA representative or open a ticket with the subject line "MCP Early Access".
• An API token associated with your Organization ID. Power My Analytics provisions this for you during alpha onboarding (you do not need to generate one yourself). Headless integrations use this token directly as a Bearer credential; browser-based AI clients use it indirectly through the OAuth flow. Rotate it from Hub settings when needed.

Requirements on the AI Client Side

Requirements for Headless / Server-to-Server Integrations

Headless integrations (custom backend services with no browser, no interactive user session) authenticate directly with a Bearer token, skipping the OAuth handshake entirely. To set one up you will need:

• An MCP client library that supports the Streamable HTTP transport with a custom Authorization header. The official @modelcontextprotocol/sdk package ships with a compatible StreamableHTTPClientTransport; equivalent libraries exist in other languages.
• Your existing PMA API token, which the MCP server accepts directly as a Bearer credential.
• A secrets-management approach for the token (environment variables, a secrets manager, or equivalent).

For the full walkthrough, see Headless/Server-to-Server Integration with PMA MCP Server.

Current Limitations

Because the MCP server is currently in Alpha, several behaviors are in flight or being tracked for improvement. Knowing them up front prevents surprises.

• Alpha access is allow-listed. If your organization is not yet on the access list, the connection will not authorize. Contact your PMA representative.
• Token lifetimes vary by authentication path. OAuth-issued access tokens (used by browser-based AI clients) last 1 year; when one expires or is revoked, disconnect and reconnect to mint a new one. The underlying PMA API token used by 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. Refresh tokens are not currently issued.
• ChatGPT requires explicit per-chat tool selection. The PMA MCP must be enabled from the tools menu in each new chat.
• Sync cadence is plan-dependent. The MCP server queries PMA's warehoused layer, not live platform APIs. Data syncs daily by default; hourly refreshes are available on Custom plans with the optional hourly refresh feature. Each account's last_sync timestamp is queryable via pma_list_data_sources for current freshness.
• Vague prompts cost more. When you do not specify metric, date range, and intent, the AI client may make several discovery calls before answering. Specify these up front to keep token spend down on the AI client side.
• Rate limit. 30 requests per minute per organization is a hard limit. Agentic AI clients in chatty loops can hit this and receive a generic "too many requests" error; this is intentional, not a bug.
• One API token per organization. If you have access to multiple hubs, you'll need to disconnect and reconnect to switch between them; there is no multi-hub toggle within a single session.
• Zero-value days are reported as gaps, not anomalies. pma_detect_anomalies explicitly excludes zero-value days from anomaly detection and reports them separately in the data_gaps field. If you ask "why didn't it flag the day I had zero spend," this is intentional.

Connect to Power My Analytics from Claude Desktop

The Claude Desktop flow is the most common way to use the PMA MCP server today. Individual Claude users on a qualifying plan can add the connector themselves; in organization contexts, the Claude organization Owner adds it once for the whole org, and individual users then connect their own PMA accounts.

Step 1: Add the Custom MCP Connector

1. Open Claude Desktop.
2. Click Settings, then select Connectors.
3. Click Add Custom MCP.
4. Paste the PMA MCP server URL: https://pma-mcp.web.app.
5. Click Save to register the connector for your Claude organization.

Step 2: Connect Your PMA Account (Each User)

1. In Claude Desktop, open Settings > Connectors and click Connect next to Power My Analytics.
2. Claude opens the PMA OAuth flow in your default browser.
3. Sign in to your PMA Hub using your existing credentials. Both Sign in with Google and email and password are supported.
4. Select the hub you want Claude to access. If you only have access to one hub, this step may be skipped automatically.
5. Wait for the "Authorization Successful" page to load.
6. Click Continue. (The button is disabled until the OAuth callback completes, so clicking too early is no longer possible; the page also auto-redirects after a few seconds, so clicking is optional.)
7. Return to Claude Desktop. The PMA MCP now appears as a connected tool, and Claude will use it automatically when you ask data-related questions.

You can revoke the connection at any time from Settings > Connectors in Claude Desktop. Your PMA Hub data is unaffected by disconnecting; Claude simply loses query access until you reconnect.

Connect to Power My Analytics from Claude Web

Claude Web (claude.ai) supports custom MCP connectors using the same setup pattern as Claude Desktop. The OAuth flow, browser handshake, and Continue-button behavior are identical; only the entry point in Claude differs.

1. In Claude Web, open Settings > Connectors. Individual users on a qualifying plan can add the connector themselves; in organization contexts, the Claude organization Owner adds it once for the whole org.
2. Add a custom MCP and paste the PMA MCP server URL: https://pma-mcp.web.app.
3. Click Connect, then complete the PMA OAuth flow exactly as described in Step 2 of the Claude Desktop walkthrough above (sign in with your PMA credentials, select your hub, and click Continue when the "Authorization Successful" page loads).

Claude Web then uses the PMA MCP automatically when you ask data-related questions.

Connect to Power My Analytics from Claude Code

Claude Code is Anthropic's command-line interface for Claude. It connects to the PMA MCP server through a one-time CLI command, then completes the OAuth handshake the same way the other Claude clients do.

Step 1: Register the PMA MCP Server

In a terminal where Claude Code is already installed, run:

claude mcp add pma-mcp --transport http https://pma-mcp.web.app/

This registers the PMA MCP server with your Claude Code configuration.

Step 2: Authorize Your PMA Account

1. Open the Claude Code CLI by running claude.
2. Type /mcp to view the list of registered MCP servers. The PMA MCP appears as unauthenticated.
3. Select pma-mcp. Claude Code displays a one-time authorization link.
4. Open the link in your browser. The PMA OAuth flow runs the same way as in Claude Desktop: sign in with your PMA credentials (Google SSO or email + password), select your hub, and click Continue when the "Authorization Successful" page loads (the page also auto-redirects after a few seconds).
5. Return to the Claude Code CLI. The PMA MCP now appears as authorized, and Claude will use its tools automatically when you ask data-related questions.

Connect to Power My Analytics from ChatGPT

The ChatGPT flow is similar but has one important behavioral difference: the PMA MCP must be selected explicitly in each new chat for ChatGPT to call it.

Step 1: Add the Custom Connector

1. Open ChatGPT and go to your organization settings (Owner role required).
2. Add a new custom MCP connector and paste the PMA MCP server URL: https://pma-mcp.web.app.
3. Save the connector for your ChatGPT organization.

Step 2: Authenticate Your PMA Account

1. From ChatGPT, initiate the connection to the PMA MCP. ChatGPT opens the same PMA OAuth flow described in the Claude Desktop steps.
2. Sign in to your PMA Hub, select your hub if prompted, and click Continue when the "Authorization Successful" page loads.

Step 3: Enable the PMA MCP in Each New Chat

Unlike Claude Desktop, ChatGPT does not invoke a custom MCP automatically. In every new chat, open the tools menu and select PMA MCP before sending your first prompt; otherwise ChatGPT will answer without using your PMA data.

1. Start a new chat in ChatGPT.
2. Click the tools or apps menu and select PMA MCP.
3. Send your prompt. ChatGPT will now use the PMA MCP tools when answering data questions.

Connect to Power My Analytics from a Headless / Server-to-Server Client

Headless / server-to-server integrations skip the OAuth handshake used by browser-based AI clients and authenticate directly with your PMA API token as a Bearer credential. This is the right path when you are building a custom backend service (for example, a Node.js API on Fly.io, a scheduled job on AWS, or an agentic system on your own infrastructure) and there is no browser available to complete OAuth.

Step 1: Retrieve Your PMA API Token

Locate the API token shared with you during your MCP Alpha onboarding (typically delivered via your alpha access support ticket) and store it in your service's secrets manager or environment-variable system. Do not commit the token to source control or expose it in client-side code; treat it like any production credential.

Step 2: Configure Your MCP Client

Configure your MCP client library to call https://pma-mcp.web.app/mcp over Streamable HTTP with header Authorization: Bearer <your-pma-api-token> on every request. The example below uses the official @modelcontextprotocol/sdk package for Node.js and TypeScript:

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";

const transport = new StreamableHTTPClientTransport(
  new URL("https://pma-mcp.web.app/mcp"),
  { requestInit: { headers: { Authorization: `Bearer ${process.env.PMA_API_TOKEN}` } } }
);

const client = new Client({ name: "my-service", version: "1.0.0" }, { capabilities: {} });
await client.connect(transport);

Any MCP client library that supports the Streamable HTTP transport plus a static Bearer token will work; equivalent libraries exist in other languages.

Step 3: Verify the Connection

Call a low-cost diagnostic tool such as pma_list_data_sources to confirm the connection is wired correctly. If the call returns the expected list of connected platforms (with their has_data and last_sync flags), your headless integration is working.

For the complete walkthrough (including environment-variable setup details, full error-handling patterns, token rotation guidance, and headless-specific troubleshooting), see Headless/Server-to-Server Integration with PMA MCP Server.

Tools Reference

The PMA MCP server exposes 22 first-party tools. You will rarely call them by name; your AI assistant picks the right tool based on the question. The reference below is mainly useful when you want to understand what the AI is doing on your behalf, debug an unexpected response, or write more efficient prompts.

Tools are grouped into five families.

Discovery and Diagnostics Tools

These tools answer "what do I have, and is it healthy?". The AI assistant typically calls one or more of these before doing analytical work.

Sub-Account and Activity Tools

Analytics Tools

These are the tools that do the actual analytical work. The AI assistant chooses among them based on the kind of question you asked.

Dataset (Data Builder) Tools

These tools let your AI assistant work with PMA Data Builder datasets: saved, blended, multi-source reporting workspaces that live in your hub.

When the AI assistant tries to create a dataset on your behalf, you will see a confirmation prompt asking you to approve the dataset name. This is intentional: datasets are persistent, and the confirmation step prevents accidental creation.

Template Browsing Tools

Basic Example: Build a Cross-Platform Dataset

This is the kind of task that takes ten minutes of clicking through the Hub UI but a single sentence in the MCP server. We will build the same thing both ways: a dataset that blends Facebook Ads spend with Google Ads spend so we can see total paid-search-and-social spend by week.

The MCP Way (in Claude Desktop)

Your prompt to Claude:

"Create a Power My Analytics dataset called 'Paid Spend Weekly' that blends Facebook Ads and Google Ads. Include spend, impressions, and clicks for the last 30 days. Then show me the blended weekly totals."

What Claude does behind the scenes:

1. Calls pma_list_data_sources to confirm both Facebook Ads and Google Ads are connected and have data.
2. Calls pma_describe_platform for each connector to confirm that the requested fields exist on the relevant report types: spend, impressions, and clicks on Facebook Ads' campaign_insights, and cost_micros, impressions, and clicks on Google Ads' campaigns_insights (Google Ads uses the plural report-type name and cost_micros in place of spend).
3. Calls pma_create_dataset with name "Paid Spend Weekly". You see a confirmation prompt; you click Confirm to approve.
4. Calls pma_add_data_table twice (once for Facebook Ads, once for Google Ads) with the date range "Last 30 days" and the three fields.
5. Calls pma_get_dataset_data with the new dataset ID and your last-30-days range.
6. Bins the rows into ISO weeks and presents a tidy table.

What you see in chat (excerpt):

I created the dataset Paid Spend Weekly in your hub and pulled the last 30 days. Here are the blended weekly totals across Facebook Ads and Google Ads:

Week starting	Spend	Impressions	Clicks
2026-03-30	$8,420.55	612,803	9,144
2026-04-06	$9,107.10	655,201	9,802
2026-04-13	$8,946.22	643,012	9,640
2026-04-20	$9,415.78	671,495	10,218

Spend is up roughly 12% from the first to the most recent full week. The dataset is now saved in your Hub at Data Builder > Paid Spend Weekly. Want me to add a third source or change the date range?

Total elapsed time: roughly 30 seconds, from a single prompt.

The Hub UI Way

The same workflow done by hand in the Hub. The result is identical (same dataset, same data, same place in your hub) but you do the clicks.

1. Sign in to your hub at hub.powermyanalytics.com.
2. In the left navigation, click Data Builder.
3. Click + New Dataset.
4. Either choose the "Facebook Ads + Google Ads performance overview" dataset template (if you want a pre-configured field list) or choose Blank.
5. Name the dataset Paid Spend Weekly and click Create.
6. Click + Add Data Table. Choose Facebook Ads, select your account and sub-account(s), choose the campaign_insights report type, select the fields spend, impressions, clicks, set the date range to Last 30 days, and save.
7. Click + Add Data Table again and repeat for Google Ads, choosing the campaigns_insights report type (Google Ads uses the plural form) and selecting the fields cost_micros, impressions, and clicks.
8. Click Run to populate the dataset. View the blended results in the dataset preview pane.
9. To get the same weekly totals shown above, you would either group by week in your destination tool (Looker Studio, Sheets, Excel) or build a small follow-on dataset that aggregates by ISO week.

The dataset created by the MCP path and the dataset created by the Hub UI path are the same kind of object. You can edit either one in the other surface. The MCP server is a different way to drive Data Builder, not a separate copy of it.

Advanced Example: Diagnose a Sudden Spend Drop

Where the MCP server really shines is when a question requires synthesizing several different angles. Imagine you log in on a Monday morning and notice that last week's ad spend looks lower than usual. You want to know:

1. Is spend actually down compared to the prior week, or does it just look low?
2. If it is down, is the drop a real performance change or are some days simply missing data?
3. If it is real, which campaigns or platforms are responsible?
4. Were any of the affected accounts' OAuth tokens recently revoked, which could explain the gap?

In a traditional reporting workflow, that is four separate Looker Studio scorecards, a Manage Accounts > History scan, and twenty minutes of manual digging. With the MCP server, it is one prompt.

Your Prompt to Claude

"My total ad spend looks down this past week vs the week before. Can you (a) confirm with a week-over-week comparison across Facebook Ads and Google Ads, (b) check whether any of the days with low spend are actually data gaps rather than real drops, (c) if it's a real drop, show me the top campaigns where spend fell, and (d) tell me whether any of those accounts have OAuth issues?"

What Claude Does Behind the Scenes

1. Calls pma_list_data_sources with connector_type_filter="facebook" and again with connector_type_filter="google" to gather the connected ad accounts.
2. Calls pma_get_date_range_comparison twice in parallel (once for Facebook Ads, once for Google Ads) with comparison_mode="wow" and metric="spend" to confirm direction and magnitude.
3. Calls pma_detect_anomalies for each platform with metrics=["spend"], sensitivity="medium", and a 14-day window via date_from, date_to, and lookback_days. The data_gaps field separates "no data reported" days from genuine spend drops.
4. For platforms where the drop is real, calls pma_query_custom with dimensions=["campaign_name"], metrics=["spend"], order_by="spend", and the affected date range to surface the campaigns most responsible.
5. Calls pma_get_token_health_summary to check whether any of those accounts are flagged as expired_needs_refresh or invalid_revoked.
6. Synthesizes the results into a single answer.

What You See in Chat

Here is what I found across your two ad platforms for the week of April 20-26 compared with April 13-19:

Headline week-over-week change:

• Facebook Ads: spend down 18.4% (from $4,902 to $3,999).
• Google Ads: spend down 4.1% (from $4,205 to $4,033). Within normal variance.
• Combined spend down 11.8%.

Real drop or data gap? Anomaly detection found:

• Facebook Ads: 1 data gap on April 24 (no data reported, possible sync gap, not a real drop). 2 days flagged as anomalous spend drops on April 22 and April 25 (z-scores -2.4 and -2.1 against the 30-day baseline).
• Google Ads: 0 anomalies, 0 gaps.

Where the real drop happened (Facebook Ads, April 22 and April 25): Top 3 campaigns by absolute spend decrease vs the prior week:

1. Q2-Awareness-USA: down $385 (campaign moved to PAUSED on April 22, per its status field).
2. Retargeting-Cart-Abandoners: down $214 (daily budget was reduced from $80 to $40 on April 21).
3. Brand-Lookalike-1pct: down $112 (CPC trended up; impressions roughly flat).

OAuth health: All Facebook Ads and Google Ads accounts in your hub have valid tokens. The April 24 gap is not explained by an OAuth issue; it most likely reflects a brief Facebook Marketing API sync delay. I recommend kicking off a manual backfill from Sources > Facebook Ads > Actions > Backfill data range for April 24 if the gap persists.

Bottom line: the headline drop is real and is concentrated in two paused/budget-reduced Facebook Ads campaigns, not a cross-platform performance issue. April 24 is a data gap to backfill, not a true zero day.

Five tools, six tool calls, one synthesized answer. The same investigation done by hand requires touching at least three different surfaces in the Hub plus manual cross-referencing.

Cost-control tip: Notice how the prompt above is precise: it names the platforms, the comparison window, and what to check. That precision is what makes this kind of investigation cheap on the AI client side. Vague prompts ("look at my data and tell me what's wrong") force the AI to call discovery tools repeatedly before it can begin, which costs more tokens and takes longer.

Troubleshooting

The fixes below cover the situations PMA's support team sees most often during the alpha. If your issue is not listed here, contact PMA support and include the AI client you are using, the prompt that triggered the problem, and any error message you received.

Connection and Authentication

• Symptom: "Token expired" or sudden loss of tool access.
  
• Cause: Access tokens are valid for 1 year, but they can expire earlier if your hub access is revoked or your PMA login changes.
  
• Solution: Disconnect and reconnect to mint a new token.
  
• Symptom: Connection fails entirely with no clear reason and your organization is not yet on the early-access list.
  
• Cause: Alpha access is allow-listed per organization.
  
• Solution: Open a ticket with the subject line "MCP Early Access" at support.powermyanalytics.com.
  

ChatGPT-Specific

• Symptom: ChatGPT answers your data question without calling any PMA tools; the answer is generic or hallucinated.
  
• Cause: Custom MCPs are not invoked automatically in ChatGPT.
  
• Solution: Open the tools menu in your chat and explicitly select PMA MCP, then re-send your prompt.
  

Tool Output and Data

• Symptom: Anomaly detection flags days where data was actually missing as performance drops.
  
• Cause: Pre-fix anomaly detection treated zero-value days as outliers.
  
• Solution: Resolved. As of the April 21 release, gaps are reported in a separate data_gaps field, and the AI assistant will surface gaps as "no data reported" rather than as drops.
  
• Symptom: A query returns no rows for a platform you know is connected.
  
• Cause: Either no data has synced yet for that platform, or the AI assistant is querying the wrong report type.
  
• Solution: Ask the AI to call pma_inspect_org_data for that connector to confirm what report types have data, then retry the query against the correct report type.
  
• Symptom: Two users ask the AI the same question and get answers in different shapes.
  
• Cause: Canonical response shapes for high-frequency questions are still being formalized.
  
• Solution: Restate the prompt with a specific output shape ("answer as a Markdown table with columns X, Y, Z") if consistency matters for your workflow.
  
• Symptom: The AI assistant is making lots of discovery calls before answering, which feels slow or expensive.
  
• Cause: Vague prompts.
  
• Solution: State the metric, date range, and analytic intent up front. Compare: "look at my data" (vague) vs "compare last-30-days spend on Facebook Ads vs Google Ads, broken down by campaign" (precise).
  
• Symptom: A tool returns no results because the connector name in your prompt does not match a connected source.
  
• Cause: Connector names must be passed exactly as they appear in your hub.
  
• Solution: Copy the connector name exactly from your connected sources list, or ask the AI assistant to call pma_list_data_sources first to retrieve the canonical name.
  
• Symptom: A newly added data source returns empty results.
  
• Cause: The initial sync has not yet completed for the new connection.
  
• Solution: Wait approximately one hour after connecting for the first sync to populate, then retry the query.
  
• Symptom: A tool call times out without returning data.
  
• Cause: Transient backend slowness or load.
  
• Solution: Retry the same query a few minutes later.
  

Headless / Server-to-Server

• Symptom: A 401 (unauthorized) or 403 (organization not authorized) error from the MCP endpoint.
  
• Cause: Missing or malformed Authorization header, or the organization is not on the alpha access list.
  
• Solution: Confirm the header is exactly Authorization: Bearer <token> (the literal word Bearer followed by a single space, then the token). If the header is correct and the request still fails, confirm your organization has been allow-listed for MCP Alpha access. For more headless-specific troubleshooting (token rotation, transport issues, rate limits in agentic loops), see Headless/Server-to-Server Integration with PMA MCP Server.
  

Disconnect and Revoke

• To disconnect from Claude Desktop, open Settings > Connectors and click Disconnect next to Power My Analytics. To remove the connector entirely, click Remove.
• To disconnect from ChatGPT, open your organization's connector settings and remove the PMA MCP custom connector.
• Disconnecting from your AI client revokes that client's OAuth token for your hub. Your PMA Hub data is unaffected. To rotate the underlying PMA API token, follow How to Generate a New API Key.

Resources and 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

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