What is MCP (Model Context Protocol)? Definition for Data Teams

MCP (Model Context Protocol) is an open protocol by Anthropic that standardizes how AI agents connect to external data sources and tools. For data teams, it means any MCP-compatible agent (Claude, Cursor, ChatGPT) can query your governed metrics through a single URL endpoint, no custom integrations required. Think of it as a universal adapter between AI and your data layer. Before MCP, every AI tool needed its own connector, its own auth flow, and its own query format. MCP collapses all of that into one typed interface. The protocol has gained adoption across major AI platforms since its 2024 release, making it the de facto standard for agent-to-data communication in production environments.

How does MCP work?

MCP uses a client-server architecture built on JSON-RPC 2.0. The AI agent acts as a client, sending structured requests to an MCP server that exposes tools and resources. Each tool has a typed schema describing its inputs and outputs. The agent discovers available tools, calls them with parameters, and receives structured responses.

The protocol is stateless per request and transport-agnostic. It works over HTTP, stdio, or WebSockets. The server defines what the agent can do; the agent decides when to call each tool based on the user's question. In practice, this means your MCP server publishes a manifest of available operations. An agent like Claude reads that manifest, understands the input types for each tool, and calls the right one. No prompt engineering required to teach the agent your API shape. The schema is the documentation.

Why do data teams need MCP?

Before MCP, connecting an AI agent to your data meant building a custom API integration for each agent. Claude needed one adapter, ChatGPT another, internal tools a third. Every integration duplicated access control logic and metric definitions. For a B2B SaaS team shipping analytics to 200 customers, that meant maintaining three copies of your tenant isolation logic across three different integration surfaces.

MCP replaces that with a single protocol. Define your metrics once in a semantic layer, expose them through MCP tools, and every compatible agent gets governed access. Multi-tenant support comes built in: publishable keys scope each connection to a specific customer's data. No raw SQL generation against unprotected tables. When a new AI tool adds MCP support, your data is already accessible with zero additional integration work.

How Bonnard uses MCP

Bonnard deploys as an MCP server with bon mcp. Each tenant gets a publishable key that scopes their queries to their data. The server exposes four tools:

• explore_schema: Lists available cubes, measures, and dimensions the agent can query.
• query: Executes a structured metric query with measures, dimensions, and filters.
• sql_query: Runs governed SQL against the semantic layer (not raw tables).
• describe_field: Returns metadata for a specific measure or dimension.

Agents discover these tools automatically. They query defined metrics, not raw tables, so they can't hallucinate aggregations or bypass access control. For example, when a user asks Claude "what was revenue last month?", the agent calls explore_schema to find orders.total_revenue, then calls query with a time filter. The semantic layer generates the correct SQL, applies tenant scoping, and returns a governed result. Read more about agentic analytics.

Related terms

• Semantic Layer
• Text-to-SQL