Anthropic released a significant update to Claude Code on desktop. The headlining feature is Previews: the ability to launch a dev server and see a running web application directly inside the Claude Code interface. Alongside that, the update adds automated code review with inline comments and background monitoring of GitHub CI pipelines with auto-fix capabilities.

Live App Previews

Claude Code's desktop app can now start a dev server and render the running application in an embedded preview pane. This means frontend developers can watch their app update in real time as Claude makes changes, without switching to a browser window.

The preview is not just a static iframe. Claude actively monitors the running UI, reads console logs, catches errors, and iterates on fixes automatically. If a component throws a runtime error or a layout breaks visually, Claude can detect the problem and attempt a correction in the same session.

Users can also select visual elements directly in the preview and pass feedback to Claude. Click on a button that looks wrong, describe what you want changed, and Claude makes the edit. This closes the feedback loop between "seeing the problem" and "fixing the problem" into a single interaction.

Automated Code Review

The second feature is a "Review code" button that lets Claude examine local diffs before pushing. Claude reads through the changed files and leaves inline comments directly in the desktop diff view, highlighting bugs, suggestions, and potential issues.

Once the review is complete, developers can ask Claude to address its own comments automatically. This creates a self-review loop: write code, get feedback, apply fixes, all without leaving the editor. For solo developers or small teams without a dedicated reviewer, this adds a meaningful quality check before code reaches the remote repository.

Background CI Monitoring and Auto-Fix

For GitHub-hosted projects, Claude Code now tracks pull request status, including CI check results. When a CI pipeline fails, Claude can attempt to resolve the failure automatically. Enable auto-merge, and Claude will merge the PR once all checks pass.

The practical workflow: open a PR, move on to a different task, and let Claude handle CI in the background. By the time you circle back, the original PR is either ready for final review or already merged. This is particularly useful for flaky tests or minor linting failures that do not require human judgment.

Session Continuity

The update also improves session portability. Running /desktop brings a terminal session's full context into the desktop application. A "Continue with Claude Code on the web" button moves local sessions to the cloud, letting developers start on desktop and pick up later from a browser or mobile device.

This continuity matters because coding sessions accumulate context. Being able to move a session between devices without losing that context means work does not get stranded on a single machine.

What This Changes for Frontend Development

The preview feature specifically addresses a long-standing friction point with AI coding assistants and frontend work. Until now, visual UI development with Claude Code required manual feedback cycles: make a change, switch to the browser, inspect the result, describe the problem in text, switch back. Each round trip added latency and required the developer to translate visual problems into language.

With live previews, Claude can see what the developer sees. The feedback loop shrinks from minutes to seconds. Combined with the ability to click on elements and describe issues in context, this makes AI-assisted frontend development substantially more practical than the terminal-only workflow.

The Bigger Pattern

This update fits a broader trend in AI-assisted development tools: the gradual absorption of adjacent workflow steps. Coding agents started by editing files. Then they learned to run tests. Then to commit and push. Now they preview UIs, review diffs, monitor CI, and merge PRs. Each step reduces the surface area of tasks that require manual developer intervention.

The question is no longer whether AI can write code. It is whether AI can own the full lifecycle of a change, from first edit to merged pull request, with the developer serving as a reviewer rather than an operator. Features like Previews move meaningfully in that direction.