Why Claude Needs a Data Warehouse Behind It, Not Just an API Connector
There are now two ways to give Claude, ChatGPT or Gemini access to your business data. The first is a direct connector: an MCP server that calls your accounting or CRM system's API at the moment you ask a question. Xero, HubSpot, Pipedrive and Intuit all publish one. The second is a data warehouse with an MCP server on top: your systems sync into a database on a schedule, and the AI assistant queries the database. Tugger works this way. This article explains when each approach fits, and why for reporting the warehouse wins.
The short version: a direct connector is good at telling you what is in a system right now. A warehouse is good at answering questions about your business. Those sound similar and are not.
What a direct connector does well
If you want to look something up, or do something, a direct connector is the right tool. "What stage is the Acme deal at?" "Create a follow-up task for Friday." "What's our bank balance today?" The connector asks the system, the system answers, and because it is talking to the live API it can also write: update a record, raise an invoice, log a call.
That is useful, and nothing below argues against having one. The problems start when the question changes from "what is this" to "what has been happening".
Six things a direct connector struggles with
1. History
"How has win rate by owner changed over the last six quarters" needs every deal and every stage change over eighteen months. A direct connector has to pull all of that through the API each time you ask, which is slow, often incomplete, and runs into the rate limits every API has. A warehouse has the history already, from the first day you connected, and adds to it on every sync.
2. Questions that span systems
"Pipeline value against what we actually invoiced." "Billable hours against the deal that sold them." "Shopify orders against stock on hand in Cin7." Each of those joins two systems, and a connector to one system cannot see the other. In a warehouse they are tables sitting next to each other, and the join is one question.
3. The business rules
Ask a direct connector for net revenue by month and you will get invoice totals, because it does not know to net off credit notes and refunds, strip VAT, or exclude the intercompany account. Those rules live in your finance team's heads and spreadsheets. A warehouse is where they get written down once, so Claude applies them every time and does not need to be told.
4. The same answer for everyone
Two people asking a direct connector the same question on the same day can get different numbers, depending on how the assistant chose to filter the data that time. When the logic sits in the warehouse rather than being improvised per question, the answer is the same whoever asks, and it is the same answer the dashboard shows.
5. Read-only
A connector that can create an invoice is not one you want an analyst asking reporting questions through. Tugger's MCP server is read-only by design: Claude queries a copy of your data and cannot touch the source system, so a mistyped question has no consequences. Most direct connectors can write within the permissions of whoever connected them.
6. Load on the source system
Analytical questions hit APIs hard. "Margin by customer type by month" against a direct connector means pulling every sales line, every time. A warehouse syncs once on a schedule, and the questions run against the copy. Your accounting system never knows the questions were asked.
At a glance
| Direct connector | Warehouse with MCP | |
|---|---|---|
| Good for | Lookups and actions in one system | Reporting and analysis across systems |
| Data | Whatever the API returns at the time of asking | Full history, synced on a schedule |
| Joins across systems | No | Yes |
| Business rules | Applied by the assistant, if at all | Defined once in the warehouse |
| Can change data | Usually yes | No, read-only |
| Also feeds dashboards | No | Yes, Power BI and Tableau from the same tables |
| Setup | Sign in, done | Sign in, first sync, then done |
Use both
This is not an argument for ripping out the native connectors. Keep Xero's or HubSpot's connector for day-to-day lookups and actions. Put the warehouse behind it for anything that is a report rather than a lookup, anything that joins systems, and anything where you need the number to be right and the same for everyone. Many Tugger customers run exactly that pairing.
How Tugger does it
Tugger connects 70+ business systems, including Xero, HubSpot, Simpro, Harvest, QuickBooks, Sage 50, Pipedrive, Cin7 and Unleashed, into a managed warehouse on a schedule. A built-in, read-only MCP server exposes that warehouse to Claude, ChatGPT and Gemini, and the same tables feed Power BI and Tableau. Tugger is UK-based and ISO 27001:2022 certified, and the warehouse runs on Azure in the UK by default or in any Azure region on request.
Related guides
- What is an MCP server?
- Best MCP data connectors for Xero and HubSpot compared
- Tugger vs Xero's own Claude integration
- Pipedrive to Claude: the official connector and when you need more
- How to connect QuickBooks to Claude in the UK
- How to build a Claude dashboard
Get started
Putting a warehouse behind Claude with Tugger takes no code and no faff. Try Tugger free for 10 days or book a demo and bring the question your current connector can't answer.
