Web analytics your AI assistant can use

Zero glue code

Most analytics tools have an HTTP API. Wiring that into Claude or Cursor means writing a wrapper, designing tool schemas, handling auth, error states, and rate limits. With MCP analytics, the AI client speaks the protocol directly. One config paste and you're done.

Read AND write

MCP analytics isn't just chat-with-your-data. Your agent can also create sites, manage tags, install tracking on new domains, and update workspace state. The dashboard becomes a fallback, not the primary interface.

Fleet operations at agent speed

"Tag every site with under 100 visits last month as 'dead.'" That's minutes of click-through-each-site-and-tag-it work, or one sentence with MCP analytics. For fleet operators running 10–50+ sites, this is real time saved.

Workspace-scoped by design

Each MCP key is tied to a single workspace. Your AI agent can't escape into other workspaces, modify billing, or escalate access. Rate-limited per workspace so runaway agents can't burn your quota.

• summarize_all_sites — fleet status + needs-attention
• rank_sites_by_growth — biggest movers
• identify_dead_sites — never-installed + silent
• detect_anomalies — unusual traffic days
• explain_traffic_change — "why did X happen?"
• get_realtime — active visitors and their latest pages over five minutes
• get_conversion_report — goal rates, funnel drop-offs, and measurement gaps

Compositions over the primitives — the agent gets analyst-grade structure, not raw counts to chain itself.

• create_site + get_install_snippet
• get_event_tracking_snippet — custom events
• get_badge_snippet — "Powered by" footer
• create_goal / list / delete
• create_funnel / list / delete
• add_site_tags + filter_sites_by_tag
• list_sites

Loose limits — these are bursty during setup, near-zero after.

• get_workspace_overview
• get_site_traffic
• get_top_pages / get_top_referrers
• get_breakdown — country / browser / device / channel / AI source / city / etc.
• get_time_series
• query_events
• breakdown_by_property — custom event props

Every primitive supports stacked filters (country=US AND device=mobile…) and custom date ranges. Tier-aware daily caps: 50/day Free → 3000/day Scale.

MCP analytics connects an AI assistant to website analytics through the Model Context Protocol. Gizmo exposes tools for installing tracking, querying traffic, and measuring goals and funnels, with documented inputs and structured results.

An MCP server publishes discoverable tools with input and output schemas for compatible AI assistants. Gizmo provides hosted access using OAuth or an API key, so you can query traffic and configure tracking without maintaining your own analytics API wrapper.

Claude Desktop, Claude Code, Cursor, Codex (OpenAI's coding agent), Windsurf, Cline, Continue, and Zed all speak MCP natively as of mid-2026, plus a growing list of less-known clients. Any of these can wire up to Gizmo with a single config paste — no per-client integration work on your end.

Install tracking, verify a site, read traffic and revenue, create goals and funnels, and analyze a portfolio of sites. Use get_realtime for five-minute activity, or get_conversion_report for goal rates, funnel drop-offs, and tracking gaps. Query tools support filters and date ranges where applicable.

Each connection is workspace-scoped — your AI agent can only see and modify data inside the workspace you consented to. OAuth grants are also scope-bound: you can grant analytics:read only (the AI can read your traffic but not create or modify anything), or analytics:read + analytics:write (read plus setup operations like adding a site). You can see and revoke every connected app at /dashboard/connect/oauth — disconnecting takes effect on the next request. We rate-limit per workspace (50 query calls/day on Free, scaling to 3,000/day on Scale) so a runaway agent can't burn your quota or our infra.

You can. The question is who maintains it. Every API integration is a maintenance debt: schema drift, auth refreshes, rate-limit changes, error handling. MCP is the protocol-level answer to that — your AI client implements MCP once, and from then on every MCP-native tool just works. For an indie operator running a fleet of sites, the cost of writing and maintaining custom API integrations across multiple tools is real. MCP analytics removes that cost for the analytics layer.

No, but it's where the leverage is. Even casual users get value: 'add foo.com to my analytics' from Claude is faster than logging into a dashboard and clicking through a setup wizard. The bigger leverage compounds for users running a fleet of 10+ sites or doing daily reporting workflows — having Claude pull last week's numbers and write a summary while you're still drinking coffee.

Sign up free at gizmoanalytics.io. In your IDE's MCP settings (Cursor, Claude, Windsurf), add a new server with URL https://gizmoanalytics.io/mcp. The IDE opens a browser tab to our consent screen, you click Allow, and the connection is live. Now ask 'add foo.com to Gizmo' and it does it. Free tier includes full MCP access — 50 query calls/day, 3,000 setup calls per day. If your IDE doesn't yet support OAuth-based MCP servers, /dashboard/connect still has the legacy API-key install for every major client.