As a Cursor user since the very beginning and a feedback and bug reporter here in the forum, I thought it could be useful to share my experience of working with Claude Desktop (instead of Cursor Agents/Glass) for the last week.
I may be Cursor-biased, since I am a happy user, especially with the Agents/Glass UI since I was looking forward to such a change and relied heavily on Cursor Cloud before it.
Here it goes.
(Remember that I am talking specifically about Agents/Glass UI vs. Claude Desktop)
Claude Pros
- /btw is awesome. I use it all the time.
- Simpler model choice, less cognitive overhead whether I’m using the right model.
- Mixing mouse and keyboard shortcuts is nice and useful.
- Background tasks are managed better and used much more frequently by the model. This makes the model look like it can handle requests as they come in.
- Claude preview eval tool for validating UI work actually works.
- Multi-PR mgmt in a single session.
- Multi-worktrees in a single session, without issues or stale UI parts.
- Clearer home screen with sessions requiring actions.
Claude Cons
-
Prompt queueing is worse and strange. It waits for the running prompt, but then shoots all pending ones together.
-
Code navigation: Not yet usable as standalone editor. Cursor is good enough but this isn’t. “Open as” offers only Cursor on my system… which is funny.
- open files/search is less mature.
- No code navigation/go-to-definition (point and click or F12/ALT+F12).
-
Diff viewer:
- No link/way to open actual file in file viewer?
- Commenting is line per line, no blocks allowed.
-
Github PRs are managed outside whereas in Cursor they can be managed inside.
-
Built-in browser has no dev tools? It doesn’t really work as a browser, only intended for preview builds.
-
I have less insight into what’s happening and how long each step took. Tested the different thinking/summary/etc views without much success.
-
I like the Plan mode better in Cursor:
- In Cursor I can still chat and read answers while the plan is updated on the side, whereas in Claude I haven’t found a way to, for example, ask about the questions it asks me when I cannot answer without discussing further (/btw doesn’t work because it doesn’t have the intermediate state yet). This is something that I need all the time to clarify stuff before making planning decisions.
- You can choose a large model for planning and a smaller one for building the plan
-
Background tasks cannot be opened/inspected
-
Forking defaults to new worktree instead of existing one. Creating a second session in the same worktree is cumbersome.
A notable difference, which may be a pro or a con is that Claude seems to work autonomously much longer with the same model. Why is that? Is it really the harness or tooling enabling the same model to go further? Is it simply more proactive and deciding itself how to move on? Or is it actually much slower and taking longer to get to the same end result? Is it fooling me and simply waiting for a compute slot? I could compare side by side a few ambitious prompts…but it’s non-deterministic anyway.
And one has to mention it, but I won’t say it’s a pro or con because it wouldn’t be fair: Claude subscription clearly offers more tokens for the buck. Cursor can’t fix that since they have to pay for Anthropic models at API pricing and can’t subsidize it the same way. However the answer is their own model Composer with its own separate credits, which in its 2.5 version really works and may be close to Sonnet or GPT-medium models.