Skip to main content
Warden Warden is a governed data-access runtime for AI agents. It runs as a server between your agents and your database, holds the credentials the agents never see, and decides what each agent may read, propose, and commit.
Warden is a binary you run, not a package you import. That is the product’s central design decision, not an implementation detail — see Is Warden a Go package?. Your agent points at Warden instead of at the database; its code barely changes.

What you can install today

warden

The gateway. Parses, classifies, governs, and audits every operation, and holds the database credentials.

warden-mcp

An MCP server so any MCP-speaking agent (Claude Code, Claude Desktop) reaches the gateway as a tool.

wardencert

The certification harness. Proves a connector’s guarantees against a real database rather than asserting them.

Container image

A distroless image running as a non-root user, plus a compose file that brings up Warden and PostgreSQL together.

Why it exists

Give an agent a database connection and you have handed it a permanent, unscoped, unaudited capability. Warden replaces that connection with a governed one:

The honest part

Databases differ in what they can enforce. Every guarantee is labelled with who actually provides it — native (the engine), enforced (Warden, by rewriting the statement), advisory (observed, not prevented), or unavailable — and a connection whose engine cannot meet its policy refuses to start. See the capability ladder.
Status. PostgreSQL is the reference connector, MongoDB the second engine, and Neo4j the third; all three are certified against a real server on every commit. MySQL is designed for but not built. Nothing here describes an engine that does not work — where a feature is planned, the page says so.

Quickstart

A governed PostgreSQL in about five minutes, with Docker.

Everything on offer

Every artefact, what it is for, and how it is distributed.