Warden, in practice.
Warden checks connected AI requests against rules you write. Request analysis runs on the gateway’s host; allowed requests continue to the configured AI provider.
Connect a tool.
The desktop app runs the gateway and its admin console. Install Warden for macOS or Windows, then open This device to connect Claude Code, Codex or OpenCode through their prompt hooks. Other clients can use the OpenAI-compatible proxy when they support a custom base URL.
Warden only checks traffic sent through a connection. Host deadlines, attachment support and failure settings affect coverage. Hooks block when the gateway cannot answer. Administrators can opt out with WARDEN_FAIL_CLOSED=0.
Write, review, activate.
Describe a restriction in Rules, review the proposed policy and activate it. Drafting does not activate a rule. Request decisions can allow, block or hold for human review; Activity records the result and its reason.
Models configures the rule writer and request judge separately. A remote rule writer or signed-in CLI can receive the administrator’s instruction and directory context. Policy analysis stays on the gateway.
Walk through a rule · Model configurationKnow what stays local.
The proxy extracts supported documents locally and holds incomplete extraction for review. The audit chain stores prompt hashes and decision metadata. Masked prompt text lives in a separate store with seven-day retention by default; set WARDEN_PROMPT_RETENTION_DAYS=0 to disable it.
Local checks do not make the destination AI provider local. Warden can miss attacks or refuse legitimate requests; model measurements apply to the evaluated dataset and configuration.
Security and retention · Model measurementsRun it on your terms.
Warden is free and open source under Apache-2.0. Published installers cover macOS Apple Silicon, macOS Intel and Windows x64. Linux users can run from source; release assets determine which builds are available.
For source setup, use Node 22.17 or newer and the pinned pnpm version. Run pnpm install, pnpm run setup, then pnpm run dev. Setup downloads local models. On shared hosts require an admin key with WARDEN_ADMIN_REQUIRE_KEY=1 and follow the deployment requirements below.