AI gets access. You keep control.

Human-governed PostgreSQL access for people and coding clients. Set the scope. Review consequential changes. Keep a record of what happened.

Open your workspace Try the access model

Alex RiveraHuman owner

Human permissionsales.ordersRead + propose writes

Codexquery.read

PostgreSQLIn scope

Example catalog. No query is executed.

PostgreSQL access. Accountable to people.Scroll to follow a request

Your tools.
One governed path.

CodexClaude CodeCursorHTTP MCPPostgreSQL

Useful access.
Nothing beyond it.

A coding client belongs to a person. Its permissions can be narrower than theirs, never wider. Try the boundary.

Interactive exampleNo live database calls

Alex RiveraHuman owner

Database: Staging
Allowed table: sales.orders
Access: Read + propose writes

Client token. Not a database password.

ScopeIn scope

Codexquery.read

SELECT id, total, status
FROM sales.orders
LIMIT 20;

Human grantClient grantAllowed

This read is inside both the human permission and the client permission.

Read-only transactionExample only

Workspace membership and database access are separate decisions. An invitation grants no database permissions.

A proposal is not
permission to run.

Give agents room to help without handing them the final decision on a production change.

Schema changes still run through Liquibase

ReviewHuman gate

Illustrative workflow1 / 3

The SQL comes first.

A client proposes an exact statement. A write preview keeps the SQL, scope, policy and expected impact together.

sales.ordersProduction

UPDATE sales.orders
SET status = 'reviewed'
WHERE id = 1842;

Requester: Alex Rivera via Codex
Expected direct rows: 1 in this example

Proposed, not executed

The access is scoped.
The accountability isn't.

A database action should come with an answer to who, what and why. Client SQL keeps that context attached.

Credentials stay server-side.

Encrypted connection credentials go only to allowlisted runners through the broker. Browsers and coding clients use separate application credentials.

Every client has a human owner.

A client cannot exceed its owner's active permissions. Revoke one client, or revoke a source grant and stop the access inherited from it.

A decision has a record.

Trace governed activity through the human, client, target, statement, policy and result. Uncertain outcomes remain visible.

EvidenceContext retained

Example audit record

A query, with its context.

Illustrative read request

Human: Alex Rivera
Coding client: Codex
Target: Staging / PostgreSQL
Resource: sales.orders
Decision: Within read scope

Inspect the example statement

SELECT id, total, status
FROM sales.orders
LIMIT 20;

Interactive example. No query was sent or executed.

No database password in the client

These are product mechanisms, not a compliance certification. Read the security model

A few clear answers.

No magic. No mystery permissions.

Read the access guide

Clear limits

Is this another SQL editor?

Client SQL is a PostgreSQL control plane. People can explore and query data, but the core job is governing who may act, which client they use, what needs review and what evidence remains.

Which coding clients can I connect?

Codex, Claude Code and Cursor connect over HTTP MCP. Each client gets a person-owned token and an explicit scope. Naming a supported client does not imply a partnership or endorsement.

Can an agent approve its own production change?

No. Production changes require an authorized independent human reviewer. A coding client does not receive review authority, and a client grant cannot expand its owner's access.

Does Client SQL replace Liquibase?

No. Liquibase owns schema migration execution and changeset history. Client SQL adds identity, access, preview evidence, review decisions and linked outcomes. The ordinary query runner does not execute schema migrations.

Is the hosted product ready for my production database?

The current implementation is being prepared for a hosted pilot. Cloud hardening and production acceptance remain open. Start with a disposable or staging database after your operator verifies the deployment. Do not treat this page as production-readiness approval.

Give access.
Keep the last word.

Start with a person, a database and a deliberate permission.

Open your workspace