Connect Snowflake and Expertise can answer questions straight from your data warehouse: find the right database and table, inspect columns, run read-only SQL on your own warehouse, watch credit spend and storage growth, and — on Cortex-enabled accounts — use Snowflake’s AI functions, Cortex Search, and Cortex Analyst. The connection runs as a dedicated read-only user in your Snowflake account: it can only ever read what you grant it, and it cannot write, ever.

What you can ask

“What databases and tables do we have in Snowflake?”
“How many orders did we book last quarter, by region?”
“Which market segment has the most customers? Exact numbers, please.”
“How many Snowflake credits did we burn this week, per warehouse?”
“What were our most expensive queries in the last 30 days?”

Connecting

Setup takes a few minutes and one copy-paste — no passwords and no key files.
  1. Open Connections in the dashboard, click Snowflake → Connect, and paste the web address from your browser while you’re logged into Snowflake (an account identifier like MYORG-MYACCOUNT works too). Add the warehouse the agent may use.
  2. Click Generate setup script. Expertise creates a key pair on the spot — the private key goes straight into our secret store and is never shown to anyone — and gives you a short SQL script with the public key already embedded.
  3. Run the script in a Snowflake worksheet (the dialog has an Open Snowflake worksheets button). It is a single statement, so any Run button runs all of it, and it works unchanged for regular databases and shares. It creates a read-only role and a key-based service user — nothing else in your account is touched. Not the Snowflake admin? Send the script to whoever is; that’s the only step that needs one.
  4. Back in the dialog, click Test connection. You’ll see how many databases the agent can now read.
You can disconnect from the Connections page any time (this deletes the stored key), and revoke from the Snowflake side instantly with DROP USER <the service user from your script>;.

Tools

Every tool is read-only — there are no write tools, by design.

Good to know

  • The agent can never write. Three independent layers enforce it: the role your script creates holds SELECT-only grants (Snowflake itself refuses writes), Expertise rejects any non-read statement before it is sent, and there simply are no write tools.
  • Runaway queries can’t run up your bill. Every query carries a statement timeout (up to 5 minutes), results are row- and size-capped, and metadata reads (listing, describing) use no warehouse compute at all.
  • Cost tracking is opt-in. snowflake_usage_and_costs needs the optional GRANT IMPORTED PRIVILEGES ON DATABASE SNOWFLAKE line that ships commented in your setup script — uncomment it to enable spend reporting.
  • Cortex features follow your account. AI functions, Cortex Search, and Cortex Analyst work where your Snowflake account has Cortex enabled and the services/semantic models exist; otherwise the agent reports Snowflake’s own message rather than guessing.
  • You control exposure with grants. The agent sees exactly what the read-only role is granted — extend or narrow access any time in Snowflake.

Troubleshooting

The setup script didn’t finish in Snowflake — its final ALTER USER line is what registers the key. Re-run the whole script (it’s safe to run repeatedly) and test again. If you regenerated the script after a disconnect, only the newest script’s key is valid.
The read-only role can’t see that object. Check the name, or grant access in Snowflake — the data-access block in your setup script shows the exact grants to repeat for another database.
The error lists the warehouses the agent can use. If the list is empty, the script’s GRANT USAGE ON WAREHOUSE line didn’t run or names a warehouse that doesn’t exist — fix the name from Back to account details and re-run the script.
Uncomment and run the script’s optional GRANT IMPORTED PRIVILEGES ON DATABASE SNOWFLAKE line — cost data lives in Snowflake’s ACCOUNT_USAGE share, which is a separate, deliberate grant.