AI assistants
DataGlass exposes a secure Model Context Protocol (MCP) server for your connected shops. It gives compatible AI assistants a structured way to analyze normalized Shopee, Lazada, and TikTok Shop data — and, when you approve a write scope, to act on it. Instead of pasting exports into a chat, you can ask a business question and let the assistant discover the authorized shop, retrieve the smallest useful dataset, and explain the result with the right financial context.
The connector is designed for decisions, not raw data dumping. Its tools distinguish realized profit from payout, headline ROAS from True ROAS, and explicit COGS from inferred costs. Eleven native playbooks describe reliable tool sequences for common investigations such as a profit decline, weak ads, low-margin products, and inventory planning.
Connect once. Ask in plain language.
Add the production MCP endpoint to your assistant, sign in with DataGlass, and choose what it may access. A new connector is read-only until you approve a write scope.
https://prod-api.dataglasslabs.com/mcpThe surface is 568 tools across three marketplaces, 263 of which can write. Nothing writes until you grant that domain a scope.
Try an interactive example
Choose a marketplace and business goal below. The lab writes a precise prompt, shows the tools an assistant should call in order, and generates an illustrative answer. It runs entirely in your browser and never connects to your DataGlass account.
Interactive example
Build a better question—and see how MCP answers it
Choose a business goal to generate a precise prompt, preview the tool route, and run an illustrative answer. No account data or credentials are used in this demo.
Summarize profit, GMV, margin, COGS coverage, and urgent signals.
Generated prompt
ShopeeGive Bangkok Home Studio on Shopee a health check for Last 30 days, compared with the immediately preceding equal-length period. Start with a one-sentence conclusion, then show realized profit, GMV, margin, COGS coverage, and loss-makers. Rank no more than three risks, state the exact dates and currency, and label any inferred COGS.
Expected tool route
- 1
list_shopee_shopsResolve an authorized shop ID and region instead of guessing.
Waiting - 2
diagnose_shopRead KPI deltas, COGS coverage, loss-makers, and the strongest signals.
Waiting
Illustrative answer
Illustrative only · no account data usedRun the example to reveal the analysis structure a good assistant should return.
How a question becomes an answer
- Resolve scope. The assistant discovers the shop identifier and region available to the signed-in account. It never guesses a shop ID.
- Start with a summary. Broad questions begin with a compact health, financial, campaign, or inventory result rather than downloading every row.
- Follow evidence. The assistant drills into product, cost, campaign, order, or settlement detail only when a signal supports it.
- Explain the decision. A good answer leads with the conclusion, shows the most important numbers, states dates and currency, and labels uncertainty.
- Ask before acting. If the next step is a write, the assistant names the tool and the exact change, and waits for your go-ahead. Your assistant prompts you too, because DataGlass flags every mutating tool as destructive.
Choose an AI assistant guide
What you can ask a DataGlass assistant
- Diagnose performance: explain profit, GMV, margin, or conversion changes and rank their causes.
- Evaluate products and ads: find loss-makers, inspect True ROAS, and compare recommended targets or budgets.
- Simulate decisions: test price, discount, or COGS changes without touching a marketplace listing.
- Run operations: review orders, unshipped backlog, stock alerts, reorder plans, forecasts, and inbound purchase orders.
- Reconcile money: distinguish realized profit from marketplace payout and investigate settlement differences.
The connector can combine these domains in one investigation. For example, a profit-drop question can begin with a shop diagnostic, identify ads cost as a material signal, verify cancellation-adjusted True ROAS, and then retrieve a target-ROAS recommendation. Each additional call should answer a specific follow-up question; it should not be used merely to produce a longer response.
What a DataGlass assistant can change, once you allow it
A connector is issued read-only. Write ability comes only from a per-domain OAuth scope you approve on the consent screen, so you can let an agent manage ad spend without letting it touch prices. With the matching scope granted, an assistant can apply price-tier drafts and target margins, create or pause ads optimization groups, write COGS and stock, act on DataGlass Actions, handle orders, returns, and shipping, and log purchase orders.
Writes go through DataGlass's own mediated flows — the ones that compute the resulting margin and record a rollback state — never through a raw marketplace passthrough. One honest consequence: vouchers, bundle deals, flexi-combo, and free-shipping campaigns have no mediated equivalent, so an assistant cannot create or edit them at any scope.
Ask AI from any screen
DataGlass writes the prompt for you. Screens that carry a full financial picture — Actions, product Financial Breakdown, the new-product calculator, the product overview, and ad campaign detail — show a small ChatGPT and Claude button. Click one and that assistant opens with the prompt already filled in; the same prompt lands on your clipboard, and Copy prompt takes it anywhere else.
The prompt asks the assistant to pull live data through the connector first, and embeds the on-screen numbers as a fallback so it still works before you connect. It follows your workspace language, so a Thai workspace produces a Thai prompt with tool names left as-is.