The Technique — Native Inline Review for Claude Code
The vscode-agent-reviewer extension creates a seamless feedback loop between you and Claude Code. Instead of copying code blocks into chat or typing verbose prompts about line 42, you can now:
- Let Claude Code modify files in your project
- Review changes directly in VS Code with inline comments
- Click "Send Review" to push structured feedback to Claude
- Watch Claude address each comment specifically
The extension auto-provisions everything: MCP server, slash command (/review), and cleanup hooks. No manual configuration required.
Why It Works — Structured Feedback via Channels API
This works because it leverages Claude Code's Channels API (currently in research preview). Channels allow external tools to communicate directly with a running Claude Code session using structured data formats.
When you add a comment on line 15, the extension packages it with:
- The exact file path
- Line numbers and surrounding context
- Your specific feedback
- Metadata about the current session
Claude receives this as a structured review object, not a free-text prompt. This means Claude understands exactly what needs fixing and where, reducing ambiguity and token waste.
How To Apply It — Setup in 2 Minutes
Prerequisites:
- Claude Code v2.1.80+
- A
claude.ailogin (API keys don't support channels yet)
Installation:
# Install the extension
code --install-extension ETSD.agent-code-reviewer
# Start Claude Code with channels enabled
claude --dangerously-load-development-channels server:code-review
The --dangerously-load-development-channels flag is required while channels remain in research preview. The extension handles registering the code-review MCP server automatically.
Workflow:
- Start Claude Code with the flag above
- Look for your session in VS Code's status bar (bottom-right)
- Click the
+icon in any line gutter to add a comment - Type feedback and press checkmark (or Ctrl+Enter)
- Click the Send icon in the editor title bar
- Select target session if multiple are running
Claude will process your structured review and respond addressing each comment individually. Click the Trash icon to discard pending comments without sending.
Limitations and Considerations
- Channels are experimental: This relies on the Channels API which is in research preview. The
--dangerously-load-development-channelsflag requirement will disappear once Anthropic approves the channel via marketplace review ororgAllowedChannelPlugins. - claude.ai login only: API key and Anthropic Console authentication don't support channels yet.
- Session discovery: The extension discovers all alive sessions via health checks and lets you pick which one to send to.
When This Beats Traditional Review
This extension shines when:
- Claude makes multiple changes across files and you need to provide targeted feedback
- You want to ensure Claude addresses every specific concern (structured format helps)
- You're reviewing complex logic where line-level precision matters
- You're collaborating with Claude on a large refactor and need iterative feedback
For simple changes, traditional prompting might still be faster. But for substantial modifications, this inline review system could cut your feedback loop from minutes to seconds.
