Claude Code updates desktop app with side-by-side sessions and integrated terminal

• Anthropic's launch post turned Claude Code on desktop into a multi-session workspace, with several Claude runs visible side by side and a sidebar to manage them.
• In the product thread, Anthropic said the desktop app now includes an integrated terminal, file editing, HTML and PDF preview, a faster diff viewer, a drag-and-drop layout, and full CLI plugin compatibility.
• The official Claude Code product page now describes desktop as a beta surface for managing parallel tasks, reviewing visual diffs, previewing servers, and monitoring PR status from one place, while Cat Wu added that the UI also shows git status and lets users pin active sessions.
• ClaudeCodeLog's release summary and the official v2.1.108 release notes show the same rollout also shipped a 1-hour prompt cache option, a new /recap command, and built-in slash commands that the model can invoke through the Skill tool.
• According to the changelog thread, 2.1.108 also reduced memory use by loading language grammars on demand and fixed an auto mode bug where the Agent tool could ask for permission after the safety classifier overflowed its context window.

You can browse the desktop product page, grab the app from Anthropic's download page, read the full 2.1.108 release notes, and inspect an external prompt diff for 2.1.108. The interesting bit is that Anthropic shipped two different layers at once: a more IDE-like desktop shell, and a CLI release that changes caching, handoff, built-ins, and some model behavior.

Anthropic's headline change was simple: one desktop window can now host multiple Claude Code sessions at once, instead of forcing the old one-terminal-per-session juggling.

On the official Claude Code product page, Anthropic frames desktop as the place to manage multiple parallel tasks, review visual diffs, preview servers, and monitor PR status. In Cat Wu's note after using it daily, the company added a few details not spelled out in the launch copy, including git status in the UI, pinned sessions, and managing both local and cloud sessions from the same surface.

The second change is that Claude Code on desktop now looks much less like a thin wrapper around the CLI.

Anthropic's launch thread names six concrete additions:

• Integrated terminal
• File editing
• HTML preview
• PDF preview
• Faster diff viewer
• Drag-and-drop layout

In TestingCatalog's walkthrough, those panels show up as Terminal, Plan, Tasks, and Preview. Anthropic also says CLI plugins work exactly as they do on the command line, which matters because it keeps the desktop app from becoming a separate tooling fork.

The same day, Anthropic published the official 2.1.108 release, and the changelog is unusually relevant to the desktop story because several changes affect long-running or resumed sessions.

The biggest changes in the release notes are:

• ENABLE_PROMPT_CACHING_1H, which opts API key, Bedrock, Vertex, and Foundry users into a 1-hour prompt cache TTL
• FORCE_PROMPT_CACHING_5M, which forces a 5-minute TTL
• /recap, which summarizes context when returning to a session
• A startup warning when prompt caching has been disabled by environment variables
• A /resume picker that defaults to sessions from the current directory
• /undo as an alias for /rewind

According to the fuller changelog list, 2.1.108 also fixed a telemetry-related bug that could knock subscribers back to a 5-minute cache TTL, fixed silent errors in --resume and --teleport, and fixed an auto mode case where the Agent tool asked for permission after the safety classifier transcript outgrew its context window.

Anthropic also expanded what Claude Code can trigger on its own.

The release says the model can now discover and invoke built-in slash commands such as /init, /review, and /security-review through the Skill tool. ClaudeCodeLog's prompt analysis adds the rest of the visible built-ins in this release:

• init, for creating CLAUDE.md
• statusline, for statusline setup
• review, for PR review
• security-review, for current-branch security review
• insights, for session reporting
• onboarding, for team onboarding guides

That same prompt diff shows Anthropic widening claude-api skill routing toward Managed Agents and file-level feature tuning, including caching, thinking, and tool use.

One more odd detail came from the additional 2.1.108 changes thread: the shipping artifact got slightly smaller while the prompt layer grew.

According to that metadata dump, 2.1.108 cut bundle size by 132.1 kB and estimated prettified LOC by 1,083, while adding 98 prompt files and 16,663 prompt tokens. The token mix also shifted a bit toward system-reminder and tools text. In the same thread, ClaudeCodeLog says Anthropic tightened pacing rules so long leading sleep commands are blocked, and polling is supposed to use a Monitor-style until <check>; do sleep 2; done loop with notification on exit.