What it is
Code Review Workbench (CRW) is a desktop app that sits between a developer and the "submit PR" button. Point it at your codebase and your team's coding guidelines, and it reviews your changes against your team's specific standards -- catching the same issues a senior engineer would flag, seconds after you save the file instead of days after you open the pull request.
CRW runs entirely on your machine. Nothing leaves the laptop except the specific text you choose to send to an AI provider (Anthropic or OpenAI). And it never edits your code or your guidelines file: every change is a proposal you approve manually. The agent can read, search, and reason about your repo, but it has no tool for writing files. By design.
Why I'm building it
Every team has a guidelines document that the generic AI tools don't know about. Linters don't read it. Cloud-based review platforms can't see it (and you wouldn't want them to). So the same drift gets caught in human code review, week after week, by the same senior engineers who could be doing something more interesting. CRW closes that loop locally: your standards, applied to your code, with no source ever leaving the machine unless you explicitly send a snippet to a provider.
The "agent that cannot write" constraint is intentional. Once an AI can edit your code without your eyes on it, you have a different product -- and a different threat model. CRW stays a workbench, not a robot.
Current state
Active development. Core architecture is set: local repo access, guideline ingestion, agentic review pass, proposal UI. AI provider integrations (Anthropic, OpenAI) wired up with the agent constrained to read-only tools.
What's next
(More to share once the next slice is solid.)