Code review for every edit your agent makes

Hooks for Claude Code, Codex and Cursor. A 3B model checks every JavaScript and TypeScript edit before it's written.

Model d1-3B Code Review Decisions

Try it

Free Hosted Endpoint

CR-1Payroll decoder
1 / 5
0 lines
Sent to the model, never stored.

Hooks for the agents you already use

Claude CodeCodexCursor

0.2 seconds per edit. 13 questions in one GPU call. Your rules decide what happens next.

Time per edit, GPU on
0.21 s
Questions per edit
13
Test score, macro F1 (max 1.0)
0.917
Cost while the GPU is off
$0
Privacy

Only the changed code

Changed lines plus nearby context. Never the whole repo.

Control

Off, flag or block

Set each of the 10 rules. Blocked edits go back to the agent with the reason.

Cost

$0 when idle

The GPU shuts off after 60 idle seconds. A slow review never holds up the agent.

Rules

Off ignores it. Flag adds a note. Block sends the edit back.

Saved into the install command below.

Threshold: the minimum model score (0 to 1) for a rule to count. Higher means fewer, surer reports.

    Install

    One command. Needs Node.js 18+ and backs up any config it touches.

    Free for a limited time, US only. The hosted review endpoint has no uptime guarantee and can be shut down at any time without notice. If it goes away, your edits go through unreviewed.

    What it sets up

      Change settings later

      crh status                  current rules
crh off | on                pause all checks
crh set unsafe_input block  one rule
crh set all flag            every rule
crh threshold 0.6           all rules
crh set unsafe_input 0.9    one rule's threshold
crh check src/api.ts        review a file now
crh update                  latest hook, same settings
crh uninstall claude        this agent only
crh uninstall all           every agent, every file

      crh lives in ~/.code-review-hooks. CODE_REVIEW_HOOKS=off skips a session.

      How it works

      The hook runs locally with your agent. The model runs on a GPU that's off when idle.

      1. 01

        The agent edits a file

        The hook takes the changed lines and your last prompt (.js, .jsx, .ts, .tsx).

      2. 02

        The model reviews it

        d1-3B answers 13 questions in one call.

      3. 03

        Your rules decide

        Over the threshold, Block sends it back and Flag adds a note.

      4. 04

        Failures never block

        If the service is down or slow, the edit goes through. Errors go to hook.log.