A year ago, “ask your analytics a question in plain English” meant a vendor’s built-in chatbot. Now it means connecting your analytics to the AI chat you already use, Claude, ChatGPT or anything else that speaks the Model Context Protocol (MCP), and asking there.
Almost every web analytics tool has an MCP server by now. They are not the same thing. Some only read reports, some can create dashboards and delete feature flags, and one of the biggest only runs on your laptop. And the hard part is rarely the connection. It’s whether the numbers that come back are right.
What We Compared
We looked at the official MCP server of each tool, or the community servers where there is no official one, from their documentation and directory listings in November 2026:
- Where it runs: hosted, or a process on your own machine.
- What it reads: reports only, or funnels, journeys and raw queries too.
- What it can change: nothing, or settings, goals, dashboards and more.
- What happens before a change: a preview you approve, a permission setting, or nothing.
The full table, with sources and cells we could not confirm, is on our AI-ready analytics page. Here is the short version.
| Tool | Official server | Reads | Can change | Before a change |
|---|---|---|---|---|
| Google Analytics 4 | Yes, runs locally | Data API reports | Nothing | Not applicable |
| Matomo | Yes, a plugin (2026) | Reporting API | Goals and segments, if an admin enables them | Admin turns write tools on; no per-change preview |
| PostHog | Yes, hosted | Insights and SQL | Dashboards, feature flags and more, including deletes | Up to your AI client |
| Rybbit | Yes, hosted (paid plans) | Stats, sessions, events, SQL | Sites, goals, funnels, members | Scoped API keys; no preview documented |
| Umami | Yes (@umami/mcp) | Traffic stats | Nothing | Not applicable |
| Plausible, Fathom | Community servers only | Depends on the server | Depends on the server | Depends on the server |
| d8a.tech | Yes, hosted | Tables, funnels, journeys, saved insights | Properties, goals, insights, traffic filters, AI context | A full preview, saved only after you approve |
GA4: Free, Read-Only, and Local
Google’s own Google Analytics MCP server is open source and, in Google’s words, “available for read requests only”. It runs as a process on your machine with your Google credentials, so it works in Claude Desktop and Claude Code but not in Claude on the web or on your phone, as Count’s setup guide points out.
People who use it like the price and the safety. The complaints are about the answers:
- Numbers that look right and aren’t. Several guides argue the API is usually correct and the mistakes happen in the model’s reading: picking
totalUserswhen you meantactiveUsers, or doing its own sums. Sequel describes the most common failure as a wrong date range reported with full confidence, and cites a Reddit report of an assistant announcing a 35% increase while traffic fell 60%. We could not verify that report, but the pattern matches what the guides describe. - Thresholding nobody mentions. GA4 hides small dimension combinations for privacy. The API response flags it, but the assistant rarely tells you unless you ask, according to the Count guide.
- Cost. In a test by Coupler.io, a single question on a large model used up a whole Claude Pro quota. Coupler sells a competing connector, so weigh that, but large unnested reports are expensive for any model.
Not everyone is unhappy. A developer writing on Habr connected the GA4 server to Claude Desktop so that Claude would “bring me all the information” instead of the author clicking through the interface, and concluded that with a good prompt it can be “a good marketer for your product” (our translation).
Matomo: Official, Careful, and Read-First
Matomo shipped its MCP Server plugin in 2026, for Matomo Cloud and On-Premise. It is built around reading reports, with optional write tools for things like goals and segments that an admin switches on one by one, capped by a maximum privilege level.
Matomo’s own security FAQ is unusually frank: it rates a read-only setup as low risk and write access as medium risk, because text stored in your analytics, such as a page title, could try to steer the assistant. Its advice is a dedicated Matomo user for MCP with the fewest permissions that work. We think that advice applies to every tool in this list.
PostHog: The Most Complete, With the Most Rope
PostHog’s hosted MCP server does the most: insights, SQL, dashboards, feature flags, experiments and more, with an AI wizard that installs the SDK and the server for you, and a Claude plugin that adds skills on top of the connector. A reviewer on Trustpilot says they plug it into Claude for all their hobby projects and call it amazing.
The flip side is in PostHog’s own docs: write tools include deletes, and PostHog tells you to always review tool calls before they run, because of prompt injection. Whether a change gets reviewed is up to your AI client and your habits, not the server.
Rybbit and Umami: Small Tools, Real Servers
Rybbit has a hosted server on its paid plans that reads stats, sessions and raw events, runs read-only SQL, and can manage sites, goals and funnels. Its docs recommend a scoped API key with only read scopes if you want a read-only setup; out-of-scope calls return a 403.
Umami’s official server is deliberately small: read-only tools over the public API, never touching the database, hosted on Umami Cloud or run with npx.
Plausible and Fathom don’t publish their own servers; community ones fill the gap, with all the maintenance questions that brings.
What Actually Goes Wrong
Across the write-ups, the same three problems come back, and none of them is about MCP itself:
- The model guesses what a metric means. If the server hands it a nested schema and 200 similarly named fields, it will pick one, confidently.
- Nobody sees the query. An answer without the date range, filters and metric behind it can’t be checked.
- Writes without a checkpoint. A server that can change settings or delete things is only as safe as the client’s confirmation dialog, and prompt injection through analytics data is a documented risk.
Our Take, and Where d8a.tech Fits
We built the d8a.tech MCP server around those three problems, and we’re biased, so here is exactly what it does:
- It gives the model definitions, not guesses. Every property comes with its timezone, goals, catalog IDs and query rules, plus a data dictionary. The assistant is told to use exact IDs and never add rows up into totals; it gets backend totals instead.
- It runs real queries and says so. Tables, time series, funnels and journeys come from queries on your data, and our weekly review recipe asks the assistant to show the query behind each claim.
- Every change is a preview first. With read and write access, it can create properties, goals and insights and edit filters and AI context. Each change is shown as a complete before and after, and saved only when you approve it. New traffic filters start in Test mode and only label traffic. Collected data can’t be changed or deleted.
It works in Claude, ChatGPT and any MCP client, read-only by default, for data in the d8a.tech cloud. It doesn’t work with your own warehouse or a self-hosted install yet; there, point your own tools at your tables.
Before You Connect Any of Them
- Start read-only, with a dedicated user or a scoped key.
- Ask for the query, the date range and the filters behind every number you plan to repeat.
- Compare one answer against the tool’s own UI before you trust the rest.
- Turn on writes only when there’s a review step you actually use.
If you want to try ours, sign up for the free Cloud Dashboard, add d8a.tech from the connectors in Claude or ChatGPT, and ask for this week’s review. If a cell in our comparison is out of date, tell us and we’ll fix it.