Klyrova

Read-only AI SQL: how Klyrova keeps your database safe

Connecting an AI agent to your database is safe when the agent can only read, when that rule is enforced by the server rather than by the AI's good behaviour, and when your credentials never reach the model or the browser. This page explains how Klyrova does each of those, and what is left for you to decide.

Start freeSee pricing

A dashboard’s datasets, one opened to show the SQL it replays on refresh

The short answer

Klyrova is built for read-only AI SQL. Every query, whether the agent wrote it a second ago or a dashboard is replaying it at 6 a.m., passes a read-only check on Klyrova's server before it reaches your database. The check accepts one statement, a SELECT or a WITH, and refuses everything else. It lives in code, not in the AI's instructions, so no prompt and no clever wording can talk it out of the rule. On top of that we recommend a database login that can only read, so your database enforces the same rule a second time.

The read-only check

Every query passes the check, wherever it comes from: the agent's questions, a dashboard refreshing, someone picking a filter, paging through a table, exporting rows, or a data model being imported. The check:

  • accepts exactly one statement, and it must be a SELECT or a WITH query, with no batching;
  • refuses SELECT … INTO, which creates a table;
  • refuses statements and functions that write, wait or reach another server, including the ones shaped like a SELECT: nextval and setval, NEXT VALUE FOR, OPENROWSET, OPENQUERY, OPENDATASOURCE, dblink, WAITFOR, pg_sleep, pg_read_file, lo_import and others;
  • on SQL Server, Azure SQL and Fabric, parses the statement with Microsoft's own T-SQL parser and requires a single SELECT without INTO, rather than trusting a pattern match;
  • reads comments and quoted names for what they are, so a table genuinely called [comment] works, while an unquoted keyword hidden in a query is still refused.

A refused query never leaves Klyrova. The agent is told why in plain words and writes a query that reads instead.

Read-only at the database too

  • PostgreSQL: every query runs inside a transaction set to READ ONLY, with a statement timeout, and the transaction is rolled back afterwards. A function that tries to write is refused by PostgreSQL itself.
  • SQL Server and Azure SQL have no equivalent session mode, so the check above and a read-only login are what hold there. Use a login that is only a member of db_datareader.
  • Every engine: a query that runs longer than 75 seconds is stopped, and a result is capped in size, so one heavy question can't tie up your database or Klyrova.

The read-only database user guide has the exact SQL for PostgreSQL, SQL Server and Azure SQL Database.

Your credentials stay with Klyrova's server

  • Database passwords are encrypted with keys protected by Azure Key Vault, bound to your account, and never returned by any part of the API, not even to you.
  • Microsoft Fabric connects through your Microsoft sign-in; Klyrova keeps an encrypted token instead of a password, and only reads the items you picked.
  • The AI model never sees a password or a connection string. It asks Klyrova's server to run a query; the server holds the connection, applies the check and hands back rows.
  • Klyrova only connects to databases on the public internet. Private, loopback and link-local addresses are refused, so Klyrova can't be used to reach into a network it shouldn't. Fabric and other sign-in-based sources may only point at Microsoft's own endpoints.

Sharing moves access, never credentials

When you share a dashboard, the people you share it with see your dashboard; they don't get your connection. Its queries always run as the dashboard's owner, through the same check. Viewers can look, filter and refresh. Editors can ask the agent for changes, but only against the data sources you allowed when you shared, and they pay for their own AI use. People other than the owner never see the database's raw error messages, which can name hosts and logins.

Dashboards run in a sealed frame

Each dashboard is a page the agent writes, so Klyrova treats that page as untrusted. It runs in a sandboxed frame with its own locked-down content policy: no network access, no cookies, no access to the Klyrova app around it. The only data it receives is the rows Klyrova hands it, and it can't send them anywhere. Published dashboards add their own rules: visitors receive the page and saved rows, never the SQL, the sources or the owner. See embedded analytics.

Connected apps are read-only too

When you connect Jira, Smartsheet or a REST API, Klyrova reads the data you chose as it reads a database: live, keeping nothing of it beyond a few minutes, or, when you import a dashboard's data model, into an encrypted copy, as you choose; either way it only ever reads from the app. Jira's sign-in asks Atlassian for read access alone. Every call goes over HTTPS to an address on the public internet: Klyrova resolves each name itself and refuses private, loopback and link-local addresses, checks every redirect the same way, and never sends your key to another host. Keys, tokens and sign-ins are encrypted and never shown again. Removing the app deletes its tables and what Klyrova holds to sign in.

What is stored, and how

  • Encrypted at rest: database passwords and tokens, each dashboard query's latest results (so a dashboard opens without re-running everything), imported copies of data models, and the steps and attached files in your chat history.
  • Where: Microsoft Azure, in the United States (West US 2). Connections are HTTPS only.
  • How long: for as long as you keep it. Deleting a source, dashboard or file removes it straight away; database backups keep it for up to seven days, and logs are kept for 30 days.
  • Payments are taken by Stripe on its own page. Klyrova never sees your card details.

What the AI provider receives

Klyrova runs the AI for you, on OpenAI's models, as our privacy policy says. During a chat, the agent sends your messages, your database's schema, the query results it needs and any files you attach to OpenAI, through Klyrova's account. OpenAI doesn't use that traffic to train its models. Refreshing a dashboard, picking filters and exporting data involve no AI at all: the stored queries run against your database and the rows go to your screen.

Klyrova doesn't sell your data, use it for advertising, or use your data or chats to train AI models.

What is up to you

  • Use a read-only login that can see only the schemas Klyrova needs. It's the strongest single step you can take.
  • Choose what to connect. Personal data in your database stays your responsibility: you are its controller and Klyrova processes it on your instructions, as the privacy policy sets out.
  • Choose who to share with, and which sources editors may query.
  • Think before publishing publicly. A public link shows the report and its rows to anyone who has it.

Questions about AI and database security

Is it safe to connect an AI agent to a production database?

With Klyrova, the agent can only run single SELECT or WITH queries that pass a server-side read-only check, PostgreSQL queries also run in a read-only transaction, and long queries are stopped after 75 seconds. Connecting with a read-only login makes the database enforce the same rule a second time.

Can a prompt make the AI delete or change my data?

No. The read-only rule is enforced by Klyrova's server on every query before it reaches your database, not by instructions to the model. A query that writes is refused whatever asked for it.

Does the AI model see my database password?

No. Passwords are encrypted, held by Klyrova's server and never returned by the API. The model asks the server to run a query and gets rows back; it never sees a connection string.

Is my data used to train AI models?

No. Klyrova doesn't use your data or chats to train models, and OpenAI, the AI provider it runs on, doesn't train on this traffic. OpenAI may keep it for a limited time to detect abuse.

Where is my data stored?

In Microsoft Azure in the United States (West US 2). Stored dashboard results, imported data models and credentials are encrypted at rest.

Does Klyrova have SOC 2 or ISO 27001 certification?

Not today. This page and the privacy policy describe exactly what Klyrova does with your data; for anything they don't answer, write to support@klyrova.app.

Updated 16 November 2026

Try it on your own data

Connect a database or drop in a file, ask a question, and keep the answer as a live dashboard. Free to start, with 500 credits and no card.

Start freeSee pricing