Model Context Protocol for Analysts: What to Check

The model context protocol for analysts comes down to one question. When an AI assistant connects to your data over MCP, can it see governed definitions, or only raw tables? And can you trace what it ran? MCP is the standard that makes the connection. It does not answer either part for you.

This guide explains what MCP is and how it works. It then covers what to check before you trust an answer.

What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard that lets AI assistants connect to tools and data sources. Instead of pasting information into a chat window, you let the assistant read files, query databases and call actions through a server. What it can reach depends on what that server exposes.

Before MCP, every AI-to-data connection needed custom engineering. Ten AI tools plus ten data sources meant up to one hundred bespoke integrations. Each one needed its own maintenance. Data Science Dojo calls this the N×M problem. Data Science Dojo

MCP changes this. Any client built to the standard can talk to any server built to it. Think of it as USB-C for AI. It is one standard port.

In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. OpenAI, Block, AWS, Google, Microsoft, and Cloudflare joined as supporters. Wikipedia

Model context protocol for analysts: the problem it solves

Most writing about MCP is for developers. The model context protocol for analysts is less covered.

Analysts lose time to context fragmentation. Every time you open an AI tool, you start from zero. You re-explain your data model and your naming conventions. You paste in query results and hope the AI understands the schema.

Mixpanel's analysis notes that analysts can use natural language queries without waiting for manual data prep. MixpanelThat is the gain. You ask in plain language and the assistant queries the source.

The risk sits in the same place. If the assistant only sees raw tables, it has to guess what "revenue" or "active customer" means. It will write a query that runs and returns a number. Whether that number matches your definition is a separate matter.

How MCP works: hosts, clients and servers

The model context protocol for analysts has three parts: hosts, clients, and servers.

The host is the application the analyst uses, such as a desktop assistant. Inside the host, an MCP client keeps a connection to each server. The server is the bridge. It exposes a specific system, such as a SQL database or a CRM, to the AI model.

IBM's overview describes three capabilities of MCP servers. Tools are actions the AI can take. Resources are data the AI can read. Prompts are reusable templates. IBM

The protocol defines how these are described and called. It does not decide which data a server exposes, how it applies permissions, or whether it keeps a log. Each server author makes those choices.

A workflow: before and after MCP

This is already happening. Moody's describes using MCP to let analysts draft from live data. Moody's

Governance and audit trails: what MCP does and does not give you

Governance is where the detail matters in the model context protocol for analysts. MCP gives servers a standard way to describe what they offer. A well-built server can limit what the AI reaches and can log each call. AtScale notes that MCP can work with existing governance and access controls. AtScale That is a property of how the server is built, not of the protocol alone.

For analytics teams, the test is provenance. Can you answer "where did this number come from?" Two things decide it. The first is what the assistant can see. A server that exposes raw tables leaves the definitions to the model. A server that exposes a governed semantic layer gives it your definitions. The second is what you can inspect afterwards: the query that ran, and a log of it.

The MCP ecosystem in 2026

The ecosystem has grown fast. By late 2025, the MCP Registry had close to two thousand servers. This was a 400% increase from its launch. Monthly SDK downloads reached 97 million. MCP Blog

For the model context protocol for analysts, the useful servers connect to data systems. Google released an MCP server for Data Commons. This gives access to public datasets through natural language. Google Developers Mixpanel also built an MCP server for product analytics.

What to check before adopting MCP

MCP adoption does not require rebuilding your data stack. Start with one or two connections. Prove the workflow. Then expand.

Security needs attention. Pento's 2025 review notes that MCP was built for simplicity. Many implementations need strong auth controls. Pento Prefer servers that use OAuth 2. Check that each server sets explicit access controls.

Think about scope. Connect AI only to the data it needs. This keeps your governance model clean. It also keeps the AI's context focused.

How Veritly approaches this

Veritly is an Integrated Analysis Environment. It puts the question, the definitions and the SQL in one place. One semantic model resolves every dashboard, document and job. When an analyst asks a question in plain English, Veritly answers through that model, so "revenue" means one thing everywhere.

Veritly shows the SQL behind the answer. You can read the query, check it and rerun it. That is the trace an analyst needs, whether the question came from a person or an assistant. To learn how a shared model works, read the semantic layer explained.

Veritly is in private beta. Join the waitlist — access opens as capacity does.

Summary

The model context protocol for analysts is a standard way to connect AI assistants to data. It removes custom integration work. It does not decide whether the assistant sees governed definitions or raw tables, or whether you can trace what it ran. Ask both questions of every setup.

For the wider picture of how AI retrieves business data before answering, see RAG for analytics.

See also: how to build an AI agent and orchestrator agents.