Add Support for Claude 3 Opus

[7KTX]

Is your feature request related to a problem? Please describe. Cursor is amazing. Claude 3 is also amazing. It's distinctly more performant than GPT-4 when it comes to coding and I think it'd be highly useful to integrate it with Cursor. Currently I am using Double instead. It's great, but Cursor is far more intuitive to use. The problem is that Claude is so much better than GPT-4 that it's not worth using Cursor without it currently IMO.

Describe the solution you'd like An implementation that works with Anthropic's Claude 3 Opus :)

[ilyakamens]

Yes, please!

[vincentsch]

Completely agree with @y0usaf. Cursor is an exceptional tool, and Opus 3 is the superior model.

Let's see if Cursor integrates Opus 3. Cursor is developed by Anysphere, which is backed by the OpenAI Fund. Opus 3 comes from Anthropic, a significant competitor to Cursor's lead investor.

[clayrisser]

Please

Thank you

😊

[ccheney]

The ability to point to any LLM should be seriously discussed, as it's highly likely that more advanced models will continue to be released in the foreseeable future.

e.g.

• Anthropic's Claude Opus
• Google's Gemini 1.5 Pro/Ultra

The app should also be flexible with the length of the context window relative to what the currently active model supports:

• gpt-4 = 128k
• claude opus = 200k
• gemini 1.5 = 1M+

[riley-wood]

Just chiming in here to add to the pile of requests for this. Claude Opus performs better than GPT-4 in my testing for coding, and the context window is much-needed for large codebase analysis. Please please please add this. I'd even pay a premium for it.

[ttizze]

please!!!!!!!!!!!!!!!

[briandonahue]

Just signed up for Claude and have been blown away. Would love to try it in Cursor.

As an idea - what if your paid subscription included the ability to choose between multiple models - ChatGPT and Claude and maybe more. I would then be inclined to pay for Cursor instead of (or perhaps in addition to) multiple AI tools...

[NicolasStuff]

I reaaaaaaly want this feature!! I would pay more for the service if it could bring Opus to Custor 🙏

[cantonalex]

Pls

[ghost]

Opus! Opus! Opus!

[NitrousBGC]

This would be an incredible addition. I would be ready to pay alot more for Opus. Additionally, lots of us can't even try it out because of region restrictions.

Let's cross our fingers! 🤞

[SebastienPingal]

Switching to double...

[bitdeep]

I'll pay more for Opus, no problem, you guys can add additional layer of subscription for model like Opus.

[tendiesfortwo]

Currently I am using Double instead. It's great, but Cursor is far more intuitive to use. The problem is that Claude is so much better than GPT-4 that it's not worth using Cursor without it currently IMO.

and here i thought i was the only one using double.bot! yes, i downloaded it only because they had claude 3 opus and the code is significantly better.

their extension has less features but in the 1 week since i downloaded, their team has shipped like 6 new versions, they move quick!!

[pb376]

Just downloaded double after seeing the rec here and yeah, the code output is indeed notably better quality when you choose claude v3, thanks.

[Ischca]

The ability to use Claude3 Opus is attractive, but what is more important to me personally is that the AI has access to the codebase.

[pb376]

I don't know why cursor devs haven't responded to this over the past week, but apparently it's been possible to use Claude this whole time, i.e. via openrouter: https://forum.cursor.sh/t/support-local-llms/1099/25

[vincentsch]

@pb376 Thank you for the suggestion.

Configuring Cursor with Openrouter.ai isn't entirely straightforward. I've prepared a step-by-step guide:

https://www.vincentschmalbach.com/adding-models-to-cursor/

[NicolasStuff]

I think they want to use Opus but they can't because they have raised funds from OpenAI...

I'll also switch to double.bot for Opus

[cantonalex]

I don't know why cursor devs haven't responded to this over the past week, but apparently it's been possible to use Claude this whole time, i.e. via openrouter: https://forum.cursor.sh/t/support-local-llms/1099/25

@pb376 doesn't support codebase

[vincentsch]

@cantonalex Cursor does send code context with every request.

Here's a breakdown of Cursor's API request: https://www.vincentschmalbach.com/analyzing-cursors-gpt-4-api-request/

(Request is the same if you use Claude 3 Opus through Openrouter.ai.)

[mike-courian]

+1

[cantonalex]

@cantonalex Cursor does send code context with every request.

Here's a breakdown of Cursor's API request: https://www.vincentschmalbach.com/analyzing-cursors-gpt-4-api-request/

(Request is the same if you use Claude 3 Opus through Openrouter.ai.)

@vincentsch no it's not thats why everyone is requesting support. It works on only the open file, not the codebase.

[InsightfulFuture]

+1 seriously that would be incredible

[murnanedaniel]

+1 This should be the absolute priority @ Cursor right now.

[lukecivantos]

+1

[thewillhuang]

+1

[lc]

Please!!

[truell20]

Hello! As a stopgap measure, if you’re a Pro or Business user, you can add “claude-3-opus” as a custom model in the Settings page, and use 10 fast requests per day for free (unlimited slow, but the delay increases exponentially).

We expect to roll out a more permanent solution (incl API key users) within the next few days.

We've made some changes to our internals to make adding new model providers much quicker. Going forward, we expect to add new near-SOTA coding models within the day they release.

[vincentsch]

@truell20

That doesn't work really well:

1. Select some code range
2. Press Ctrl+K
3. Type a prompt + Enter
4. Selected code will be deleted and replaced by this sentence: "Here is the rewritten selection according to your instructions:"

Tried adding this to my prompt: "just give me the rewritten code, nothing else, no explanation, just the code"

Initially this seemed to work better. But the generated code is truncated and leaves me with a non-functioning half-finished thing.

Update: For smaller files or pieces of code it's working, but you always have to add something like "just give me the rewritten code, nothing else, no explanation, just the code"

[murnanedaniel]

I'm getting the same behaviour. Ctrl K not working as before

[transcendr]

@vincentsch thanks for the guide, appreciated it. That said using openrouter seems to break Cursor's context for me. When sending a request with documentation, in the default setup, it reads the docs and generates doc-specific code. Making the same request with openrouter enabled results in generic code and the model responding things like "since I don't have access to xyz, here's a general approach you could take" and it always sends the current file as context which the default doesn't do (and is silly to send whatever random file you have open as context with each request). There's definitely something not working quite right.

[vincentsch]

@transcendr Yes, I wasn't satisfied with how it worked either. We'll have to wait until the Cursor team fully integrates Opus, I guess.

[NicolasStuff]

Hope it's coming soon!

[vincentsch]

@NicolasStuff It's already there.

Here's a step-by-step guide for adding Claude 3 Opus to Cursor:

1. Open Cursor.

2. Hit Ctrl+Shift+J to enter Cursor settings.

3. Scroll to "OpenAI API".

4. Click on "Configure models" below the API key field.

5. Click on "+ Add model" (below the model names).

6. Enter "claude-3-opus" as the model name.

7. Click on the "+" next to the field to add the model.

8. Close the settings page.

9. Use Cursor as before with the new Claude 3 Opus model.

That's it! With the recent update, you no longer need to use OpenRouter.ai or override the OpenAI Base URL. Just add "claude-3-opus" as a custom model in the Settings page, and you're ready to go.

Edit: You have to add "just give me the code, nothing else, no explanation, just the code" to every prompt to make it work.

[clayrisser]

@vincentsch how to add the Claude api key?

I tried add it instead of the openai api key.

I get the following error.

I'm using the latest version of cursor.

[clayrisser]

I think I need to override the OpenAI Base URL. @vincentsch what should it be set to?

I tried the following.

https://api.anthropic.com/v1 https://api.anthropic.com

[briandonahue]

@clayrisser noticed this in the release notes - not sure if you are a paid user, but it doesn't seem like it's fully supported (yet):

If you’re a Pro or Business user, you can add "claude-3-opus" as a custom model in the Settings page and use 10 fast requests per day for free (unlimited slow, but the delay increases exponentially). We expect to roll out a more permanent solution (including API key users) very soon.

[vincentsch]

@clayrisser You cannot yet use your own API key. See what @truell20 said above. They will roll out support for custom API keys in the next few days.

[7KTX]

Well, looks like support is pretty much here. It looks like we can't click on "Apply" like normal, but I think it's still pretty good. Thanks a lot Cursor team!

[LifeGains]

@truell20 i do not see the Claude API key options on my end anymore - it was there yesterday? On a Mac