Claude's new Cursor killer just dropped

Concise summary

A hands-on review of Anthropic’s new Claude Code desktop app (bundled into the main Claude app with Chat and Co-work). The reviewer appreciates the GUI and lower RAM usage versus the prior CLI but finds the release buggy, half-baked, and lacking polish. They recommend open alternatives (Codeex, T3 Code) until Anthropic improves UX, QA, and openness.

Overview

• Claude Code is now available as an official desktop GUI (Chat + Co-work + Code in one app).
• The app reduces memory/CPU use compared to the prior CLI/VM service.
• Despite performance improvements, the release is criticized for many UX issues, bugs, and questionable product choices.

Key positives

• Official desktop GUI combining Chat, Co-work, and Code.
• Lower memory and CPU usage than the previous CLI approach.
• Promising feature set:
  • Split/tiled chat views and multi-chat tiling.
  • Ability to add multiple folders as context.
  • Remote-run (run sessions on other machines on your local network).
  • Routines/templated runs (cron/webhook).
  • Skills/connectors and a plugin/connector store.
  • Forking chats and a “preview” developer workflow (run dev server, inspect logs/network).
• Easier model toggles and ability to switch Opus model versions (reviewer prefers Opus 4.5 over 4.6/461m).

Major problems and UX/bugs (detailed)

• Image paste / attachment issues
  • Pasted screenshots attach to the wrong message or appear out of order.
  • UI sometimes requires a manual refresh to show attachments.
  • No easy delete UI for attachments.
• Missing convenience affordances
  • Many places lack click-to-copy buttons.
  • Copy/paste behavior can add extra newlines or wrap oddly.
• Agent behavior and permissions
  • Agents sometimes perform side effects (create branches/PRs, modify permissions) when only analysis was requested.
  • Bypass/permission preferences aren’t reliably saved.
• Hotkey, focus, and keybinding problems
  • Global hotkeys act on the wrong pane (often the leftmost window) rather than the focused pane.
  • Terminal opens in the wrong tab.
  • Command-W has edge case behavior when closing panes.
  • Several hotkeys map to suboptimal functions (e.g., command-number toggles chat/co-work/code unexpectedly).
• Split view and layout instability
  • Tiling/opening threads causes layout shifts.
  • Drag/drop ordering and sidebar customization are poorly implemented.
  • Many UI elements don’t behave as expected.
• Git / worktree handling
  • Default worktree storage can pollute repositories (users must add .gitignore entries).
  • Little visibility into the current worktree/branch state.
• Platform limitations and product choices
  • Cloud Code requires a specific project structure (quad files/quad MD) — feels like vendor lock-in.
  • Some connectors exposed may violate platform ToS.
• QA and polish issues
  • Abrupt bugs and UI oddities: menus disappearing, mismatched settings placement, duplicate voice input, diff view rendering issues.
  • Reviewer finds this level of polish surprising for a well-funded company and compares unfavorably to more mature alternatives.

Comparisons and alternatives

• Codeex / Codex (open-source CLI + app server)
  • Praised for reliability and wide adoption as a harness for agent tools.
  • Easy to integrate with other GUIs; recommended if you want an open solution.
• T3 Code (reviewer’s open-source project)
  • Presented as a more polished alternative: better add-project flow, favicon grabbing, proper multi-tab handling, hotkeys that respect focus, one-click copy, robust diff rendering, long-thread performance, drag/drop project ordering.
  • Claimed to be faster, more stable, and more usable despite being a two-person project.
• Cursor and other GUIs
  • Some competitors adopt community standards (agents directory, agent.md).
  • Anthropic’s approach remains more proprietary by comparison.

Practical tips, fixes, and workflow notes

• Model selection
  • If Opus 4.6 / 461m underperforms for you, switch to Opus 4.5 locally.
• Git / worktree
  • Expect to add .gitignore entries for Claude’s built-in worktree defaults to avoid repo pollution.
• Remote-run and Cloud Code
  • Remote-run sessions are possible on your network, but mobile use is limited.
  • Cloud Code for web is a different product and won’t run code locally.
• If you want an open foundation
  • Use Codeex’s Apache-licensed CLI/app server as a base; consider T3 Code for a GUI.

Reviewer’s verdict and analysis

• Pros: fixes some performance problems (notably RAM) and provides an official GUI.
• Cons: far behind in polish, stability, and UX. The app feels rushed, with many basic UI/QA failures.
• Strategic critique: Anthropic’s decision to keep core pieces closed/proprietary is criticized — the reviewer argues Anthropic should either open key components or deliver a world-class UI; currently it’s doing neither.
• Practical recommendation: Because Claude is a common sign-in path, the app may still see wide use, but serious coders should evaluate alternatives (Codeex-based tools, T3 Code, Cursor) until Anthropic improves the product.

Guides / demos shown in the video

• Pasting screenshots into chat and issues observed.
• Using split/tiled views and opening terminals.
• Switching Opus models.
• Adding projects and grabbing favicons.
• Installing the T3 Code nightly build.
• Showing code diffs and long-thread navigation.
• Demonstrating how worktrees are created and why .gitignore edits are needed.
• Promise of a future in-depth comparison video across multiple GUIs for agentic coding.

Sponsor

• Code Rabbit — an AI code review platform. Presented as a tool for automated code review to catch small mistakes and free humans to focus on high-level review.

Main speakers / sources

• Video narrator / reviewer: creator of T3 Code (primary speaker).
• Companies / products discussed: Anthropic (Claude / Cloud Code), Codeex / Codex, T3 Code, Cursor, OpenAI, Code Rabbit.
• Individuals referenced: Julius and Julia (T3 Code collaborators), Theo and Ben (mentioned in commentary).