Delivery

MCP server

What it is

A Model Context Protocol server, so Claude, Cursor, Codex or your own agent can read your analytics directly. One endpoint:

https://app.vitrus.dev/mcp

Two ways in. Clients that support it sign in with OAuth: you add the URL, a browser window opens, you pick a workspace and click Allow. Everything else takes a workspace API key (Workspace → API keys, admins only, shown once). Either way the connection reads one workspace, can never write, and you can cut it off: OAuth apps under Account → Connected apps, keys under API keys.

Connecting

ChatGPT (Plus, Pro, Business, Enterprise, Edu — web): Settings → Apps & Connectors → enable developer mode, then add a connector with the URL https://app.vitrus.dev/mcp and sign in.

Claude (claude.ai and the desktop app): Settings → Connectors → Add custom connector → https://app.vitrus.dev/mcp, then Connect and sign in.

Claude Code, with OAuth (then run /mcp and choose Authenticate):

claude mcp add vitrus --scope user --transport http https://app.vitrus.dev/mcp

…or with an API key, no browser step:

claude mcp add vitrus --scope user --transport http https://app.vitrus.dev/mcp   --header "Authorization: Bearer vk_YOUR_KEY"

Claude Code plugin — the server plus four skills (/vitrus:weekly-report, /vitrus:traffic-drop, /vitrus:measurement-plan, /vitrus:ai-search-audit):

claude plugin marketplace add Vitrus-Dev/vitrus
claude plugin install vitrus@vitrus

Cursor — one click: Add Vitrus to Cursor. Gemini CLI: gemini extensions install https://github.com/Vitrus-Dev/vitrus. The server is listed in the official MCP Registry as dev.vitrus/analytics.

Cursor by hand (~/.cursor/mcp.json), Windsurf, Claude Desktop and any client that takes a JSON config — leave out headers to sign in with OAuth instead:

{
  "mcpServers": {
    "vitrus": {
      "type": "http",
      "url": "https://app.vitrus.dev/mcp",
      "headers": { "Authorization": "Bearer vk_YOUR_KEY" }
    }
  }
}

Transport is plain HTTP, one JSON-RPC POST per call, protocol 2025-06-18. OAuth follows the MCP authorization spec: discovery via /.well-known/oauth-protected-resource, dynamic client registration, PKCE (S256) required, one-hour access tokens, refresh tokens that rotate on every use — a refresh token used twice revokes the connection. Removing someone from a workspace also cuts off the apps they connected. Self-hosted: the same tools ship in @vitrus/mcp; mount it behind your own auth.

Tools

Every tool is read-only and says so to the client (readOnlyHint). Ranges are days or exact from/to dates.

Tool	Returns
`list_sites`	Sites you can read.
`get_overview`	Visitors, sessions, pageviews, bounce, duration, channels, pages, referrers.
`query_stats`	Any metric (visitors, sessions, pageviews, events) by any of 20 dimensions or by day, with filters.
`get_realtime`	Who is on the site now, on which pages, from where.
`get_ai_traffic`	AI referrals and AI crawlers, kept separate.
`get_revenue`	Revenue, orders, AOV, conversion — per currency, never converted.
`list_goals` · `goal_report`	Saved goals and their conversions; ad-hoc goals too.
`analyze_funnel`	Ordered funnel; you can pass your own steps.
`get_journeys`	The most common page paths, in order.
`get_retention`	Cohort matrix, or the reason it is unavailable.
`get_web_vitals`	Core Web Vitals (p75) and slowest pages.
`get_errors`	JavaScript errors by message, page and browser.
`get_digest`	The plain-language summary with evidence ids.
`get_tracking_snippet`	The script tag, custom events, revenue and identify.

Every answer carries its evidence

This is the part that differs from other analytics MCP servers. Each metric comes back with the SQL that produced it, its parameters and the raw rows:

{
  "metric": "visitors.unique",
  "value": 40,
  "query": "SELECT COUNT(DISTINCT visitor_id) ... bot_kind = ''",
  "params": ["site-id", 1789003421000, 1789608221000],
  "rows": [...]
}

A human reading a dashboard can sense when a number looks wrong. An agent cannot — it takes the number, writes it into a report, and the error propagates with full confidence. Returning the query lets the agent cite evidence, and lets you audit what it said afterwards.

It is read-only, deliberately. An agent cannot create, delete or reconfigure anything. Analytics is a system of record; a record a model can write to stops being a record. A wrong read produces a sentence you can catch — a wrong write silently corrupts the history you measure against.