Claude Code Desktop Fixes CLI but Delivers UX Slop
Anthropic's new Claude Code desktop app beats the laggy CLI on performance but ships buggy UX, proprietary lock-in, and fewer features than open alternatives like Cursor and T3 Code—builders should skip it.
GUIs Beat CLI for Real-World Agentic Coding
Graphical interfaces solve CLI pain points that hinder agentic workflows. Pasting screenshots works seamlessly: drop an image, and it appears inline for reference, unlike CLI where formatting breaks on copy-paste. Text selection preserves structure—copy code from GUI chat, paste into editor with perfect word wrap and no extra newlines. CLI mangles this, inserting artifacts and wrapping spaces oddly. Image pasting in CLI attaches correctly as "image one" for model reference, while Claude's desktop botches it, linking to prior messages or requiring manual refresh.
Multi-project handling shines in GUIs. Hover a sidebar project, hit new chat, instantly context-switch without hunting folders. Add projects via fuzzy search better than macOS picker, pulling favicons for visual ID. Claude forces manual folder navigation from a tiny recent-projects dropdown, omitting sidebar items. Split-view tiling lets multiple threads run side-by-side (right-click or Cmd-drag), enabling parallel agent tasks—Claude's version layout-shifts horribly on resize.
Hotkeys like Ctrl-` open project-specific terminals, but Claude ignores focus: terminal spawns in wrong pane, tab-traps input, X-button overlaps grabber. "Users find edge cases... models do not," explaining AI-built UIs' happy-path focus but edge-case failures.
Claude Code's Half-Baked Execution
Performance improves: desktop uses less RAM (2.5GB vs CLI's higher draw) and avoids CLI freezes, but UX regresses. Agent runs halt mid-task, icon lingers misleadingly active. No copy buttons—manual select-paste. Resizing warps layout. Permissions reset despite "bypass always"—re-asks on edits. File-open adds useless empty entries; package.json copy half-works.
Worktrees default to ".claude" in project root, demanding gitignore tweaks—only tool forcing repo mods. Sticks to proprietary ".claude" files/folders, ignoring standards like agent.md or .agents/. Cursor embraces them. No branch/worktree visibility in threads; fork chats but hide context.
Bugs cascade: image paste sends pre-message, attaches wrong; voice mode stop ambiguous (thread or transcription?); settings button misaligned, dropdown vanishes on top-item toggle. Multi-tab Cmd-W closes right pane fine, left triggers new-chat view. "This feels like... a UI that was \h__\h out with a single prompt."
Model quirks persist: Opus 4.6 (dumber than 4.5) defaults; easy toggle now exists. Remote control for networked machines promising, but mobile ties to web-only Claude Code.
"Anthropic is failing absurdly at both open APIs or great product."
Open Alternatives Dominate with Reliability
Cursor leverages open-source Codex CLI app-server (Apache 2.0)—plug any UI, no harness needed. Powers forks without Anthropic's lock-in. T3 Code matches: project fuzzy-add, favicons, intuitive UI. Nightly adds theming. Both outfeature Claude: reliable pastes, standards compliance, no gitignore hacks.
CodeRabbit sponsor highlights AI review synergy: catches syntax while humans eye architecture. Bun, Clerk, Nvidia use it—CLI/IDE integrations build confidence, free brain for big-picture.
Anthropic's hype (official tweet, months teasing) contrasts slop. Community accepts it via Claude sub convenience—"two less clicks to sign in... 15 more to do anything." Locks users despite dozens better/open options.
Stealable: built-in worktrees (checked default), multi-folder context pre-thread, tiling. But execution fails: no QA, edge cases ignored.
"Too many of y'all just accept the \h__\h slop... Despite... dozens of better options."
"The CLI is such a trash piece of software that anything is better than it."
Key Takeaways
- Ditch Claude Code CLI/desktop for Cursor or T3 Code—open Codex CLI server enables custom UIs reliably.
- Prioritize GUI agent tools: screenshot/text paste, split-views, fuzzy project switch save hours vs terminal hacks.
- Audit git workflows—avoid tools dumping .claude folders; enforce agent.md standards.
- Test edge cases in AI UIs: focus wrong panes, resize shifts, permission resets kill productivity.
- Integrate AI code review like CodeRabbit early—frees humans for architecture, cuts bugs.
- Demand open harnesses from vendors; build on Apache 2.0 like Codex to avoid lock-in.
- Use Opus 4.5 over 4.6 locally; toggle in-app now possible.
- Fuzzy-search project adders > macOS picker; snag favicons for sidebar scanability.
- Split-test agents across tabs/projects—parallelism accelerates iteration.