I am aware from Claude’s support documents that Claude Code CLI implements more features than the Claude Code extension. I am also aware of the general differences between Claude Code CLI and the Cursor agent chat.
What is not clear to me is the difference between using Cursor’s agent chat with Anthropic models and using the Claude Code Cursor/VS Code extension. I’ve been searching for a while to find an answer, but haven’t come across any documentation.
Is anyone able to provide a little more information here? Thank you.
It’s been a minute since I used Cursor, so minor details may have changed, but generally speaking some differences include:
Using Claude Code (CC) for VS Code extension or CC CLI means you can use a CC subscription rather than the API which is much cheaper. Using an Anthropic model through Cursor means paying API prices (+markup) is your only option now that Anthropic restricts subscription oauth tokens to their harnesses only.
Different harnesses… CC extension (or CC CLI) is a different harness than Cursor uses. The Claude Code harness is quite good and works well with CC features like skills, custom agents, hooks, etc. For most of the history of AI coding agents, Anthropic has been the leader and everyone else (Cursor included) has been copying what they do. Anthropic invented MCP and was the first to implement a variety of addon agent features like (the modern version of) skills. Everyone else has been playing catch up.
Meanwhile, Cursor has been the leader in GUI and as a result their agent harness has upsides (more user friendly diff and rollback handling, better overall GUI) and downsides (token optimization strategies that lower quality, agent instructions that tell the agent not to tell you things that the CC version will happily talk about). Note that Cursor’s token saving strategies are a way of making the more expensive path (API rather than subscription) feel less painful, and they work pretty well if you don’t mind the tradeoffs.
One of the most important differences is that CC harnesses (CLI, Extensions, new desktop app) are specifically designed to work with Anthropic models, and the models are trained to use features from their harnesses fluently. Generally speaking you’ll get slightly better tool use out of the box with an Anthropic model in an Anthropic harness.
In my opinion Cursor’s biggest strength, and biggest weakness, is that it strives to be model agnostic. Well has no choice really, Composer isn’t going to compete with frontier labs on complex tasks any time soon. So you get a lot of model flexibility in trade for weaker performance from SOTA models (in some areas) than you’d get using their custom harness.
Thanks for providing your insight on this! I am looking to strike the right balance between these tools, so I appreciate your input on some of the detailed differences. I find having a large, interactive visual diff super handy, thanks tor Cursor, but I notices some relative limitations now and again (e.g. tool usage).