Review AI agent changes locally before they go in.
AI coding agents write code fast. Reviewing it is the bottleneck.
You're staring at a terminal diff, trying to parse what changed across eight files. You spot something wrong on line 47 but there's no way to point at it. So you type "actually, on line 47 of auth.ts, the token expiry should be..." and hope the agent understands which line you mean.
This is a code review problem. We solved it years ago with pull requests. But PR review UIs are designed for commits that have already been pushed. Not for local changes from an agent sitting in your terminal waiting for feedback.
The feedback loop between you and an AI agent should feel like a pull request review. You see the diff. You click a line. You leave a comment. The agent reads your comments and addresses each one.
That's what linemark does. Type /linemark in Claude Code, and it opens a diff review UI in your browser. You annotate. You submit. Structured feedback goes back to the agent.
No copy-pasting line numbers. No "on line 47 of auth.ts." Just click, comment, submit.
linemark-demo.mp4
curl -fsSL https://raw.githubusercontent.com/gdaybrice/linemark/main/install.sh | bashThis clones linemark to ~/.linemark and registers /linemark as a global Claude Code slash command. Works in any project. Just Node.js 18+, zero other dependencies.
- Make changes (or let Claude make them)
- Type
/linemark - Review the diff in your browser
- Comment on anything you want changed
- Submit to send feedback, or Approve to accept
Claude reads your comments and addresses each one. You can run /linemark again to review the next round. Repeat until you're happy.
- Side-by-side and unified diff views
- Commit selector: working tree, individual commits, or all commits
- Inline comments on lines, ranges, or whole files
- Image attachments (paste, drag, or pick)
- File tree with search
- Collapsible files with "Viewed" tracking
- Syntax highlighting
- Light and dark themes
/linemark starts a local server and opens a review UI in your browser. It defaults to showing working tree changes, and you can switch between individual commits or all commits from the dropdown. When you submit, your comments come back as structured feedback on stdout. Claude Code reads them and acts on each one.
One file for the server. One file for the UI. No build step. No dependencies beyond Node.
