Databases
Connect a PostgreSQL database (including Supabase and Neon) and ask it questions in plain English. Read-only, and only the tables you choose.
Connect a PostgreSQL database and GPTExcel can answer questions about it in a chat: it reads the structure of the tables you allow, writes the SQL, runs it and explains the result. It only reads. Every statement is checked before it runs and runs in a read-only transaction, so nothing in your database changes.
It works with any PostgreSQL server (AWS RDS, Google Cloud SQL, Azure, your own) and with hosted Postgres such as Supabase and Neon.
Plans
| Plan | Connections | Exports |
|---|---|---|
| Free | 1 | Questions only |
| Plus | 2 | Yes |
| Pro | 5 | Yes |
| Team | 5 per member | Yes |
| Business | 10 per member | Yes |
A query costs 1 credit and an export 2, on top of the chat message. A statement that's refused (out of scope, blocked, or waiting for your approval) costs nothing.
Connect a database
Open Settings → Connectors
Under Add a connector, click PostgreSQL, Supabase or Neon. They open the same form with that provider's defaults and setup steps.
Fill in the connection
Paste a connection URL into Paste a connection URL and the fields fill themselves, or type them in: Name (yours, for example "Orders database"), Host, Port, Database, User and Password.
Under TLS, keep Verify certificate (recommended). If your provider uses its own certificate authority, paste its certificate into CA certificate (optional) (see Supabase below), or choose Encrypted, not verified: the connection is still encrypted, but GPTExcel can't confirm it reached your server.
Allow our address
If your database only accepts known IP addresses, allow the address shown under Allow our address in the form (it has a copy button). GPTExcel always connects from that one fixed address.
Save and test
Click Save and test. A passing test shows Connected with the PostgreSQL version and how many tables the user can read. If something is wrong, you'll see what in plain words, with a tip; the driver's own message is behind Details. A connection that fails stays saved and is marked failing until a test passes.
Choose tables
After a passing test, click Choose tables (also in the connection's menu) to decide what the AI may use: everything except some schemas and tables, or only the tables you tick.
Use a read-only user
GPTExcel only reads, but the strongest limit is the database user itself. Copy the read-only user script in the form creates one. Run it as the owner of your tables, set a real password, then connect as that user:
-- Run as the owner of your tables, then connect GPTExcel as gptexcel_reader.
CREATE ROLE gptexcel_reader LOGIN PASSWORD 'choose-a-strong-password';
GRANT CONNECT ON DATABASE yourdb TO gptexcel_reader;
GRANT USAGE ON SCHEMA public TO gptexcel_reader;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO gptexcel_reader;
-- Tables created later are readable too:
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO gptexcel_reader;
-- Other schemas: repeat the last three statements with their name instead of public.If the test says some tables can't be read, the GRANT lines most likely ran in a different database.
Supabase
- In Supabase, click Connect and choose Session pooler. Copy the connection string and paste it into
Paste a connection URL. It fills in the host (
aws-…pooler.supabase.com), port5432and user (postgres.<project ref>); add your database password. - To verify the certificate, open Project Settings → Database → SSL configuration, click Download certificate, and paste the file's contents into CA certificate (optional). Or choose Encrypted, not verified.
Use the Session pooler, not the direct connection: Supabase's direct host (db.<project ref>.supabase.co) only
has an IPv6 address, and GPTExcel connects over IPv4.
Supabase's own schemas (auth, storage, realtime and the rest) are left out of a new connection, so the AI
only sees your tables. You can change that in Choose tables.
Neon
In Neon, open your project's Connect dialog, copy the connection string and paste it into Paste a connection URL. Every field fills in, and Verify certificate works as it is: Neon's certificates come from a public authority.
Ask questions in a chat
Attach a connection to a chat from the message box's + menu → Connect a database (up to 5 per chat). Or just ask about your database: if none is attached, GPTExcel shows a card to attach one and continue.
Then ask in plain English:
How many orders did we ship each month this year?
Which 10 customers spent the most in the last 90 days?
Compare average order value by country for Q3 and Q4.
Each step shows as a card:
- Schema: the tables and columns GPTExcel read before writing SQL.
- Query: the result as a table (up to 200 rows), with the SQL it ran. Click a card to open it.
- Export: a file in the chat, for results too big to read row by row.
GPTExcel aggregates in SQL (GROUP BY, filters, LIMIT), so answers come from small, exact results rather than
from reading whole tables.
Exports
On paid plans, GPTExcel can export a query's full result as a Parquet or CSV file in the chat. The Python sandbox reads it directly, so you can go on to charts, statistics, models or an Excel workbook, or download the file.
Export all orders from 2026 and build a workbook with revenue by month and a chart.
What keeps it safe
- Read-only. Every statement is checked before it runs; anything that writes or changes the schema is refused with a Blocked card. Allowed statements run in a read-only transaction that is rolled back.
- Only your chosen tables. Before a query runs, GPTExcel plans it (without running it) and checks every table it would read against your choice. Anything else comes back out of scope, free.
- System catalog closed. Queries can't list tables you left out through PostgreSQL's catalog
(
information_schema,pg_catalog). - You approve costly queries. If the planner expects a query to be very expensive, it waits on its card until you click Run anyway.
- Limits. Up to 30 statements a minute per person, and 3 at once per connection.
- Missing permissions show the
GRANTstatement that fixes them, and aren't retried. - Credentials. The password is stored encrypted and never shown back. See the Privacy Policy for what is stored and for how long.
Troubleshooting
| What you see | What to do |
|---|---|
| Couldn't reach the database | Check the host and port, and allow our address in your database's network settings. |
| The database didn't allow our address | Add the address under Allow our address to the database's allow list. |
| The database refused the sign-in | Check the user and password. On Supabase's pooler the user is postgres.<project ref>. |
| Couldn't verify the database's certificate | Paste your provider's CA certificate, or choose Encrypted, not verified. |
| The database needs an encrypted connection | Don't turn TLS off: keep Verify certificate or Encrypted, not verified. |
| This database doesn't exist | Check the Database field; Supabase's is postgres, Neon's is usually neondb. |
| The database is out of connections | Close idle connections or use your provider's pooler host. |
| The database gave us an outdated session | A pooler session went stale. Click Test again. |
| No permission to read this | Run the GRANT statement shown on the card as the table's owner. |
| Some tables can't be read | The GRANT lines ran in another database. Run them in the one you connected. |
Related
- Connectors: Google Drive, OneDrive and databases at a glance.
- Ask your Postgres database questions in plain English: a worked example.
- Use case: data science.