MCP server: your analytics, in your AI tools

MCP, the Model Context Protocol, is the emerging standard for letting an AI assistant use an outside tool in a structured, permissioned way. The Voris MCP server is our side of that connection. It gives your assistant a small, defined set of read-only tools that answer questions about your Voris data.

You connect it once to an MCP-compatible client, Claude Desktop, ChatGPT, Cursor, or any other, and from then on your assistant can pull your numbers into the conversation. You ask in plain language. Behind the scenes the assistant calls a scoped tool, Voris runs the same query your dashboard would, and the answer comes back grounded in real data rather than invented.

The server exposes your analytics as a handful of focused, read-only tools. In practice that means you can ask your assistant to:

• Pull your top-line metrics. Pageviews, visitors, sessions, bounce rate, and average session length over any date range, with the option to look at humans only.
• Rank what matters. Your top pages, referrers, countries, devices, campaigns, and entry pages, sorted by the metric you care about.
• Follow conversions and revenue. Goal conversions over time, attributed to their source or campaign, so you can ask where your best sign-ups actually came from.
• See who cited you. Which AI assistants cited your site, and, when you want it, what those visitors did next.
• Separate agents from people. The split of your traffic between verified AI agents, other bots, and real humans.
• Ask in plain language. Put a question to the built-in analyst and get an answer drawn from your real data, with its confidence and any caveats stated honestly rather than hidden.
• Read your tracking plan. Check what you are currently set up to measure, without leaving your assistant.

No new script, no export, no copy-paste. If your site already sends events to Voris, the MCP server is a new way to read what you have collected.

Every tool the server exposes is read-only. Your assistant can query your analytics; it cannot change a setting, delete a record, move a plan, or spend a cent. There is no destructive action to misfire, no matter how a prompt is phrased. That is a deliberate boundary, enforced on our side, not a promise about how carefully you word things.

Each connection uses a key you generate in your dashboard. The key is tied to a single site, checked on every request, and revocable the moment you want it gone. Your assistant only ever sees the site that key is for, and nothing beyond the read scope you granted. Connect one site, or issue separate keys per site, and pull any of them the instant you are done.

Connecting an AI assistant to your data should not quietly widen what that data is. So the analyst answers from pre-aggregated numbers, never from raw free-text you have stored, which means a stray value in a custom property cannot leak into a model. The queries run on our EU-hosted infrastructure, and they return the same figures you see on your dashboard, because they run the same query underneath. Same privacy line, same numbers, now reachable from the tools you think in.