Connect your own data
Give Claude a spreadsheet, keep a weekly export fresh, or connect your company's database, then ask for a report on your own tables.
Your app reads data through its connectors: one for each kind of data. A file needs no setup at all. A database needs a few connection details from whoever runs it. Either way you tell Claude, and Claude does the work.
Give Claude a file#
A CSV or Excel export of one table is the quickest start. Drag the file into the Claude chat, or say where it is on your computer, and tell Claude what it is:
“Here is orders.csv, our web shop orders since January. Put it in the CSV folder and tell me which columns it has.”
Claude copies it into the app’s folder for that kind of file:
| File | Folder |
|---|---|
Excel, .xlsx |
app/data/excel/ |
CSV, .csv |
app/data/csv/ |
JSON, .json |
app/data/json/ |
An app made with Create’s defaults reads Excel and CSV files. For JSON, ask first: “Add the JSON connector.” An older Excel file (.xls) is not read: save it as .xlsx first.
Each file becomes a table, named after the file: orders.csv is the table orders. Names are made lowercase, and spaces and other signs become _: the column Order Date is order_date. A column that holds only numbers is read as numbers.
Excel sheets#
Every sheet of an Excel file is its own table, named <file>_<sheet>: the sheet Regions of budget.xlsx is budget_regions. A sheet’s first row is its header, and a cell with a formula reads as its last calculated value. To see what a workbook gave you, ask:
“I gave you budget.xlsx. Which tables did it make, and which columns does each have?”
Refresh an export every week#
When a new export arrives, replace the old file with the new one, under the same name and in the same folder. You can do that yourself in Finder or File Explorer, or ask Claude:
“Here is this week’s orders.csv. Replace the old one and check every report still loads.”
The app notices the file changed and reads it again. A report you opened in the last ten minutes can still show the old figures, because results are kept that long so pages open fast: wait, or ask Claude to restart the app.
Keep the column names the same from week to week. A report reads columns by name, so a renamed column makes its query fail until Claude updates it.
Connect a database#
Ask IT for a read-only user#
Ask whoever runs the database for a read-only user: a login that can read the tables you need and change nothing. For example:
“I am building reports on my own computer with a tool that only reads data. Could I have a read-only database user for the orders and customers tables, and its connection details: the server’s address, the port, the database name, the user name and the password?”
Some teams send these as one connection string: a single line of text that holds everything the app needs to reach a database, in this order: the kind of database, the user name and password, the server’s address, the port and the database name. For example postgres://reports_reader:secret@db.example.com:5432/analytics. Postgres and MySQL take a connection string; SQL Server and Redshift take the same parts one by one. Connectors lists what each needs.
BigQuery signs in differently, with a service account: a Google Cloud user made for a program rather than a person. Ask whoever runs your Google Cloud project for a service account with read access to BigQuery and its key file (a .json file). Save it on your computer and give Claude its path, not its contents.
Add the connector#
If you did not tick your database in Create, ask:
“Add the Postgres connector.”
Give Claude the details#
Claude writes them into app/.env, the app’s private settings file. It stays on your computer and is never part of a saved version. Claude puts the details there and nowhere else: not in a report, a .sql file or a wiki page.
The details you type in the chat are part of the conversation, which Claude (and so Anthropic) sees. To keep the password out of it, say:
“Open app/.env for me. I will paste the password in myself; fill in the rest.”
Check it works#
“Which tables can my reports read? List the ten largest.”
If a report says Not connected, ask Claude: “This report says Not connected. Fix it.”
The app reads your data on your computer. Claude sees what it works with to write a report: table and column names, a few sample rows, query results and the files it writes. Use a read-only user, and follow your company’s rules for AI tools. See Your data and Claude.
Ask for a report on your own tables#
Ask Claude to look first, so you both use the same names:
“Look at the orders table and tell me which columns could give a weekly order count and amount.”
Then ask for the report. Say five things: its name, the table, the figures, the period and the rules for counting.
“Make a report called [name] from [table or file]. Show [the figures] per [week or month], with [tiles for the totals, a line chart, a table]. Count [the rule, such as: only paid orders, without VAT]. Follow the date filter.”
For example:
“Make a report called Orders from the orders table. Show the number of orders and their total amount per week, with tiles for the totals and a line chart. Count only paid orders, without VAT. Follow the date filter.”
Then look, check one figure against a number you trust, and steer in plain words, as in Your first report.
Next#
If something does not work, paste what you see to Claude and ask it to fix it. Then, when you want to know more: Modules, Report files, All components and FAQ.