Product thesis
A code review comment is an interruption. It asks somebody to stop, re-read their own work, and consider changing course. If that interruption comes from a machine, the bar should be higher, not lower.
That is the small bet behind Warden. The useful part of an AI reviewer is not confidence. Confidence is cheap. The useful part is the discipline around what it is allowed to be confident about.
The model is a writer, not a witness.
Warden starts with tools that can produce facts: TypeScript, ESLint, dependency audit output, duplication checks, deterministic detectors, and source snippets from the repository. Those facts are not glamorous. They are also the only reason the final review deserves attention.
The model enters after the evidence exists. Its job is to rank, connect, clarify, and phrase. It can turn a pile of diagnostics into a readable comment. It can ask a careful question when intent is unclear. It should not be treated as the source of truth for a CVE, a library API, or a claim about what the codebase does.
Local-first is not nostalgia.
The CLI is the first product surface because it is the smallest honest loop. It runs where the repository already lives. It can use the same type checker, the same lockfile, the same local cache, and the same diff a developer is about to review. There is no hosted backend pretending to understand a project before the local engine has proved it can help.
That choice also keeps the future open. A GitHub bot, a CI wrapper, or an interactive triage app can come later because the core returns a typed review contract instead of terminal-only prose. The CLI is not a dead end. It is the place where the review logic gets good before distribution becomes the problem.
Citations are part of the interface.
A citation is not decoration. It is how the reviewer says, "Here is why I am allowed to say this." Warden verifies external claims before they reach the final comment, and it drops claims that cannot be grounded. That makes the product less magical, but much more useful.
The aim is not to flood a pull request with every possible issue. The aim is to produce a small set of comments that are specific, ranked, and auditable. If Warden is unsure, it should ask. If it cannot verify, it should stay quiet.
The public surface stays honest too.
This site is static on purpose. The example output comes from a checked-in fixture. The design notes link back to the actual decision record. There is no demo box quietly calling a hosted reviewer, no fake sample repo invented to flatter the tool, and no claim that the bot or triage app already exists.
Warden is a personal tool with a public craft bar. The reason to share it is not that every future surface has shipped. The reason is that the core idea is sharp enough to inspect: an AI reviewer should bring evidence before it brings prose.