Connect Claude to Business Data via MCP: A Practical Guide

To connect Claude to business data, use a server that speaks the Model Context Protocol (MCP) and exposes your analytics, CRM or revenue data. Add it as a connector or configure it in Claude Desktop, scope access narrowly, then test with a question whose answer you already know before relying on anything it says.

What you are actually connecting

Claude on its own knows nothing about your company. MCP is the open standard that lets an assistant discover a server’s tools, such as “get analytics” or “search contacts”, and call them. The protocol’s official site documents the specification, and Anthropic announced it in a public post. For a business-oriented explanation, see BusinessMCP’s documentation on what MCP is.

You have three broad options.

Option	What it is	Fits when
Hosted analyst endpoint	A vendor exposes unified business data over MCP	You want working data quickly
Individual MCP servers per tool	One server for analytics, one for CRM, and so on	You like control and have engineering time
Custom MCP server	You build one over internal systems	You have proprietary data or strict requirements

Example: connecting the BusinessMCP endpoint

BusinessMCP documents how to connect Claude, ChatGPT, Gemini and other clients in its connect Claude, GPT and Gemini guide. According to that page:

• The endpoint is the same for all clients.
• For Claude Desktop, you get a bearer key from Settings in the dashboard and add an entry to the config file that runs the mcp-remote bridge with the endpoint and an authorization header, then restart. The page warns that a bare URL and headers entry is ignored.
• For Claude.ai and ChatGPT, you add the endpoint as a connector in the app’s interface and sign in via OAuth, with no key needed, then choose an access level on the consent screen.
• Gemini and other clients use the same URL and key; the page gives no Gemini-specific interface steps.

Follow that page for current syntax rather than copying from us, since configuration details change. The analyst described on the BusinessMCP AI business analyst page is the same one exposed through this endpoint, with scoped access policies controlling what an external agent sees.

Key takeaway: connection is the easy part. Scope, verification and approval rules are what make it safe to use.

A safe setup sequence

1. Choose one question. For example, “How many sign-ups did we get last week by channel?”
2. Connect only the sources needed. Not everything, just what answers that question.
3. Pick the narrowest access level. Read-only first. The access control documentation describes scoped policies.
4. Verify discovery. The BusinessMCP page suggests confirming tools such as get_analytics appear in the list.
5. Ask a known question. One where you can check the answer in the source tool.
6. Compare. Recompute the number. If it differs, find out why before going further.
7. Add sources and actions gradually. Each addition gets its own test.

Which data is worth exposing

Start with what a typical operator asks about. BusinessMCP’s data you can query page lists areas such as web analytics and funnels, CRM and identity graph, company enrichment, SEO, ads and social stats, and a company knowledge base, with some features limited to certain plans. Use the list as a menu, not a checklist.

Data	Value	Sensitivity
Aggregate web analytics	High	Low
Revenue and payments summary	High	Medium
Ad spend and ROAS	Medium to high	Medium
Contact-level CRM records	Situational	High (personal data)
Email and chat content	Situational	High

Start low on sensitivity and add higher-sensitivity data only with a reason and a legal basis. More on this in AI analyst data privacy and GDPR.

Troubleshooting

• Server never appears in Claude Desktop. Check the config uses the bridge command format the documentation specifies, and restart fully.
• Tools listed but answers empty. The key’s policy may redact fields or block a data class; check scope.
• Answers disagree with your dashboard. Check definitions and date windows before blaming the model.
• A write action does not happen. Some actions return a pending state and need approval in the dashboard. Per the BusinessMCP documentation, sending email requires owner approval.

Risks to keep in view

Anything an assistant reads is exposed to the model, and untrusted text can include instructions meant to manipulate it. Keep write permissions limited, require approvals for outbound actions, and log activity. BusinessMCP’s guide to MCP security best practices is a useful checklist, and our article on AI agents for business analysts covers the permission model in more depth.

After it works

Write down the five questions your team now asks the assistant weekly, who may ask them, and who reviews the answers. Re-test after any change to tracking, tools or definitions. And keep your verification habit: can you trust AI data analysis explains what to check and why.

Choosing your first three questions

A connection earns its place by answering real questions. Pick three that your team currently answers by hand every week. Good candidates are a weekly sign-up and revenue summary by channel, a list of high-intent visitors who have not converted, and a check on which campaigns are spending without producing revenue.

Write each as a short prompt with the window, the metric and the breakdown spelled out, and save it in a shared note. Run the same prompt each week so results are comparable, and keep a column for whether someone verified the numbers. After a month you will know which prompts the team relies on, which need rewording and which questions the connected data cannot answer. That list, more than the connection itself, tells you whether to add sources, tighten access or stop.

Who should have access

Connecting one person’s Claude is different from connecting a team. Decide whether each person gets their own key or sign-in, so that activity is attributable and one person’s departure does not break everyone. Give analysts broader scope than casual users, and keep write-capable access for a small group. Revoke access promptly when roles change, and review the list of active connections every quarter. The access-control model in the BusinessMCP documentation, with scoped policies and revocable keys, is a good pattern to ask any other vendor to match.