What MCP solves
Before MCP, every AI tool needed a custom integration for every external system it talked to. ChatGPT had plugins. Claude had its own tool-use format. Gemini had extensions. Each ecosystem was siloed, with vendor-specific APIs and limited interoperability.
MCP standardises that. An MCP server exposes a set of tools, resources, and prompts. Any MCP-compatible AI client can talk to it without bespoke integration work. Build one MCP server for your analytics, and Claude, ChatGPT, and Gemini can all query it.
For data and analytics, this turned out to matter a lot. Users no longer had to copy-paste numbers from a dashboard into a chat, the chat could fetch the number itself.
What MCP defines
The protocol has three primary primitives:
- Tools, actions the AI can call (e.g.
get_attribution_metrics,list_campaigns) - Resources, data the AI can read (e.g. dashboards, reports, customer lists)
- Prompts, pre-defined templates the user can invoke (e.g. “monthly performance review”)
A server publishes a manifest of what it offers. The client (AI app) discovers the manifest and decides what to call.
What MCP enables for analytics
Concrete patterns that work in 2026:
- “What’s our ROAS by channel this quarter?”, chat fetches live numbers, no dashboard required
- “Which case studies should I send this prospect?”, chat queries a CRM-aware MCP server, returns curated picks
- “Draft this week’s performance recap”, chat pulls metrics from MCP, drafts a summary, user reviews
- “Why did our Meta CAC spike yesterday?”, chat queries multiple data sources, correlates, hypothesises
The unifying pattern: the AI does the analytical legwork, the human does judgment and action.
Security in MCP
By design, MCP is permission-aware:
- Tool calls require user approval (in most clients) before execution
- Read-only access is the default pattern for analytics MCP servers
- Authentication happens at the MCP server level, typically OAuth or API key
- The user controls which servers their AI client trusts
A well-designed analytics MCP server exposes read-only tools, scopes data access to the calling user’s permissions, and logs every query for audit.
What MCP doesn’t do
- It doesn’t make AI smarter. Garbage data through MCP still produces garbage answers.
- It doesn’t replace semantic layers. AI is bad at interpreting raw schemas. A semantic layer adds the metric definitions and business logic that make MCP-returned numbers correct.
- It doesn’t fix tracking. If your underlying data is wrong, MCP delivers wrong answers efficiently.
The combination that works: clean first-party data → semantic layer for metric definitions → MCP server exposing the semantic layer → AI client.
Common mistakes
- Exposing raw warehouse tables via MCP. Too much surface area, no business context, easy to produce confidently-wrong answers.
- Skipping the semantic layer. Without consistent metric definitions, two AI queries can return contradictory numbers.
- Treating MCP as a chatbot. It’s plumbing. The user experience is the AI client. MCP is what makes the AI client useful.
FAQ about MCP (Model Context Protocol)
What is MCP (Model Context Protocol)?
MCP is an open protocol that lets AI tools like Claude, ChatGPT, and Gemini query external data sources and call external tools in a standardised way. Anthropic introduced it in late 2024 and the major AI ecosystems have all adopted it.
What does MCP do for analytics?
MCP lets users ask AI assistants questions about their real data, “what is our blended ROAS this week?”, and get live answers from the actual numbers, not guesses from the model’s training data.
Is MCP safe to use with sensitive data?
When designed correctly, yes. MCP servers are typically read-only and scoped to the calling user’s permissions. Tool calls require user approval in most clients. The architecture is permission-aware by design.
Why does MCP need a semantic layer?
Without a semantic layer, the AI sees raw warehouse columns and writes SQL that is syntactically valid but semantically wrong. A semantic layer exposes consistent metric definitions so the AI returns trustworthy numbers.