Skip to content

Journal

·6 min read

An MCP server so agents read the same numbers you do

Six read-only tools — stats, pages, sources, AI traffic, goals, insights. Claude, Cursor, and Codex can brief you from observed rollups, not from a CSV you pasted.

MCPAgents

Two computer terminals facing each other in a dark room
The agent reads what the dashboard already shows. Nothing else.

Founders already ask agents to write the Monday note. The usual path is an export, a paste, and a model that now holds a week of rows it should never have seen. Vector 5 ships a read-only MCP server instead. The agent calls the same aggregates the dashboard renders. Nothing at event level is exposed.

.mcp.json
{
  "mcpServers": {
    "vector5": {
      "command": "npx",
      "args": ["@vector5/mcp"]
    }
  }
}

The six tools

  • query_stats — visitors, pageviews, bounce, duration for a window and actor filter.
  • top_pages — paths with views, and crawl-to-referral rates where they exist.
  • sources — referrers and UTM, including assistant hosts.
  • ai_traffic — crawler fetches, human_via_ai, headless agents.
  • goals — named events with optional props, still filterable by actor.
  • insights — the latest aggregate brief, if you enabled generation.

Skills for Claude, Cursor, and Codex sit next to the server (MIT, like the tracker). They know to ask for observed numbers, to refuse to invent visitors, and to keep the cannot-know list in the same breath as the KPIs.

An agent that can read your analytics should not be able to read your visitors.