MCP for business metrics: let Claude and AI agents read your numbers without the warehouse keys
What the Model Context Protocol is, why pasting sales files into a chatbot is risky, and how to give Claude or your own agent read-only access to approved metrics.
By The DataxLand team· · 5 min read
You already use Claude or ChatGPT to draft emails and summarise contracts. The natural next step is to ask about your own business: "Which channel was most profitable this month?", "Draft the weekly report for our partner." The fastest way is to export a spreadsheet and paste it in. It works, but with three problems: raw data (including buyer details) leaves your systems; the model guesses the formulas, so numbers may not match your dashboard; and next week you paste again.
MCP (Model Context Protocol) is the alternative: instead of handing data to the AI, you give it a limited set of tools to ask for numbers when it needs them. This post explains MCP in business terms, lists the safety questions to ask, and shows how DataxLand exposes approved metrics over MCP.
MCP in one paragraph
The Model Context Protocol is an open protocol that lets AI applications (clients: Claude Desktop, Claude Code, Cursor, your own agent…) connect to external data and tools (servers). A server publishes a list of tools — name, description, parameters — and the client lets the model call them. Think of it as a standard socket: instead of every AI writing a custom integration with every system, one MCP server works with many clients.
What matters for business data: the server decides what the AI can do, not the AI. If the server only offers "read metric" tools, the AI cannot drop a table or read a customer's phone number, whatever the prompt says.
Three ways to give an AI your numbers
| Paste a file into a chatbot | Give the AI SQL access to the database | MCP through a semantic layer | |
|---|---|---|---|
| Data leaving your systems | The whole file, including personal columns | Anything a query touches | Only the aggregated numbers asked for |
| Who decides the formula | The model guesses | The model writes SQL | A versioned metric definition |
| Matches the dashboard | Not reliably | Not reliably | Yes — same definitions, same published warehouse |
| Revoking access | You cannot un-paste a file | Rotate database credentials | Revoke the token; effective on the next call |
| Audit trail | None | Database logs (if enabled) | Every call logged with token and tool |
| Fresh data | Paste again | Yes | Yes |
Five safety questions before connecting an AI to data
- Can any tool write? For Q&A, read-only should be the default. Ask: does any tool delete, modify or send anything?
- Is there a free-form SQL or query tool? If so, the AI can read everything the database user can.
- Where does personal data go? Do buyer names, phone numbers or addresses ever appear in tool results?
- Do tokens expire, and can they be revoked? A non-expiring token in a config file on a laptop is a long-term risk.
- Is there an audit log? After an incident you need to know which token called which tool, and when.
What DataxLand exposes over MCP
The DataxLand MCP server exposes the same tools Si uses — the same handlers, not a reimplementation:
| Group | Tools (examples) | Who gets them |
|---|---|---|
| Metrics | List metrics, describe a metric, query a metric by dimensions/filters/period, semantic query (several metrics, period comparison), explain a change | Any token with metric-read scope |
| Context | The workspace's approved business model | Any token with metric-read scope |
| Operations | Data health, build history — no cell values | Only technical tokens granted ops scope |
| Drafts | Propose a metric, pin a tile to a page — drafts awaiting approval only | Only tokens explicitly granted this scope |
There is no tool that runs model-written SQL, and no tool returns buyers' personal data. A token is shown once at creation and only its hash is stored; tokens expire (90 days by default) and can be revoked at any time. Every call is logged — token, user, workspace, tool, number of rows returned — without the result values.
For a token that covers several workspaces (Enterprise plan, e.g. an organization with one workspace per brand), each call must name the workspace, and the server checks the token is allowed for it before running anything.
Setup in five minutes
- In the DataxLand app: Workspace settings → MCP → Create token. Choose scopes (metric read by default), expiry and a label.
- Paste the configuration into your client, for example Claude Desktop or Cursor:
{
"mcpServers": {
"dataxland": {
"url": "https://<your MCP server>/mcp",
"headers": { "Authorization": "Bearer <MCP token>" }
}
}
}- Try: "Using DataxLand, list the workspace metrics", then "Profit after fees last month by channel".
- When you no longer need it, or suspect a leak, revoke the token on the same page.
Step-by-step details are in the MCP server docs.
What about ChatGPT and claude.ai on the web?
The DataxLand MCP server accepts two kinds of authentication. Clients that let you set headers (Claude Desktop, Claude Code, Cursor, custom agents) send a token in a header. The web apps of claude.ai and ChatGPT use the OAuth sign-in flow for MCP: add the server address, sign in to DataxLand, pick the workspace and scopes; OAuth-issued tokens also expire, can be revoked and every call is logged.
A real use: a quarterly note in Claude
On the DataxLand sample data (a simulated scented-candle shop, 1 Aug – 29 Oct 2026), a prompt like:
"Using DataxLand, write a quarterly P&L note for our partner: profit by channel, the channel to worry about, three actions."
makes Claude call the metric tools a few times and write: the shop made 89,301,200 ₫; TikTok Shop lost 13,802,100 ₫ with a 12.41% return rate; website ads reported 92,828,000 ₫ more revenue than actually arrived. The prose is Claude's; the numbers come from DataxLand, identical to the dashboard. Read it before sending — the model can still misinterpret, but it cannot invent a different definition.
MCP or Si?
- Use Si (in the app, Telegram, Slack) for quick questions without a separate AI.
- Use MCP when you already work in Claude, Cursor or another agent and want business numbers there — drafting reports, combining with other documents, or running an agent on a schedule.
Both read the same metrics, so Si and Claude never give two different numbers. More: MCP feature page · What is a semantic layer · the Shopee AI agent page.