Claude's new Cursor killer just dropped
TL;DR
Anthropic finally shipped a GUI for Claude Code, but Theo says it mostly wins by not being the CLI — he calls the old CLI “slow” and “laggy,” and says the new desktop experience is a huge improvement mainly because the baseline was so bad.
The app is packed with obvious UX and QA failures — Theo hits bugs almost immediately: frozen threads that still look active, image paste attaching to the wrong message, hotkeys targeting the wrong pane, broken split-view behavior, and permission settings that don’t persist.
His core complaint isn’t just polish, it’s platform strategy — Theo argues Anthropic is failing at both letting the community build on top of Claude Code and shipping a first-party app good enough that nobody would need to.
OpenAI’s Codex gets unexpected praise for the infrastructure layer — while Theo is still mad the Codex app itself is closed source, he highlights the Apache-licensed Codex CLI app server as a big reason so many third-party tools support it out of the box.
Theo uses T3 Code as a direct comparison and says two people built a more reliable app in a fraction of the time — he points to features like better project switching, one-click copy, functioning diff views, correct hotkey behavior, and smoother long-thread performance.
There are a few legitimately good ideas in Claude Code desktop — built-in worktrees, multi-folder context, remote machine control, and especially tiled multi-chat views stand out, but Theo says every promising feature feels “half-baked” once you touch edge cases.
The Breakdown
The Big Reveal: Claude Code Gets a Desktop App
Theo opens by dunking on the current crop of AI coding apps because they all look the same, then lands on the actual announcement: Claude Code now lives inside Anthropic’s main Claude desktop app alongside Claude chat and Claude Co-work. His top-line reaction is clear: this is a major step up from the CLI, and Anthropic is clearly treating it like a big release after months of teasing from the official Claude account.
Why a GUI Matters More Than Terminal Purists Admit
Before roasting anything, he defends the category itself. His example is simple and practical: copying formatted output and pasting screenshots are dramatically better in a GUI than in Claude Code’s terminal flow, where wrapping, newlines, and attachments get mangled. Even here, though, he finds a bug where pasted images get sent as a separate message and sometimes attach to the wrong tool call.
The Roast Starts: Bugs, Freezes, and Terrible Interaction Design
Once he starts actually using the app, the energy shifts hard. A thread freezes mid-run but still shows Claude’s icon as if it’s active, there are no copy buttons, resizing the window causes layout weirdness, and actions like “open file” produce confusing empty states. Theo’s line is basically: yes, it’s better than the CLI, but only because “the CLI is such a trash piece of software that anything is better than it.”
Codex’s Open App Server Makes Anthropic Look Even Worse
He then zooms out from UI bugs to architecture. Theo says Codex is better not just because the UX is tighter, but because OpenAI shipped an open, Apache-2 app server through the Codex CLI, which makes it easy for third-party apps to integrate without building a whole harness. That becomes his bigger frustration with Anthropic: if they won’t open up the pieces builders need, the first-party app has to be excellent, and in his view it absolutely isn’t.
The Good Ideas Buried Inside the Slop
To his credit, he does stop and call out what’s promising. Split view for multiple chats is genuinely cool, built-in worktrees are useful, multi-folder context before starting a thread is smart, and remote control for other machines on your network hints at a compelling workflow. But almost every compliment gets followed by a bug — hotkeys open terminals in the wrong pane, command-W behaves differently on the left and right side, and edge cases instantly break the illusion.
T3 Code Enters the Chat
A huge chunk of the video becomes a live side-by-side with Theo’s own app, T3 Code. He shows project switching with favicons, a faster add-project flow, one-click copy, cleaner resizing, working diff views, and correct multi-tab terminal behavior, emphasizing that many of these features took “two prompts” or were built by just two people. It’s competitive, yes, but it’s also how he makes the point that Anthropic’s misses are not hard unsolved problems.
What This Says About AI-Built Software
Theo briefly softens and says he recognizes the pattern: AI is good at building the happy path, but terrible at finding edge cases. His example is command-W in split view — it works fine on one pane, then fails bizarrely on the other — which he treats as a signature sign of software assembled without enough real user testing. That leads to his central frustration: users keep tolerating “slop,” so labs keep shipping it.
The Bigger Thesis: This App Is About Lock-In and Showcase, and It Fails at Both
Near the end, Theo argues there are only two reasons a lab builds a desktop app like this: lock users into its ecosystem, or showcase how good its models are. He thinks Claude Code desktop is trying to do both while delivering a product so buggy that it undermines Anthropic’s own claims about using it internally. The final verdict is brutal: the one clear improvement is memory usage — he sees Claude’s virtual machine service at about 2.5 GB RAM, less than before and even less than the CLI — but beyond performance, he says “everything else kind of sucks.”