Using VS Code, Cline, and Deepseek/Gemini 2.0 as Alternatives to Cursor

The cursor software is a great tool that can help beginners with no coding experience develop software from scratch. The recently popular “Kitten Light” app was developed using Cursor, which has gained widespread attention and made more ordinary people aware of the effects and potential of AI programming. Cursor is a paid software, and after a 14-day trial period, if you wish to continue using it, you need to pay $20 per month. I believe this price is worth it for a professional programmer or individual developer looking to advance in the AI field, but it might be a bit expensive for an ordinary person who occasionally develops small tools. Although there are “cracking” methods available online, a more permanent solution is needed.

Essentially, Cursor is an AI-driven code editor developed based on Visual Studio Code, integrating advanced LLMs such as GPT-4 and Claude 3.5, aiming to enhance development efficiency and code quality. Following this idea, we can look for alternatives from three aspects: code editors, large models, and tools that can utilize large models, and combine them to achieve Cursor-like functionality. Through some searches and practices, I found this combination:

• Native Visual Studio Code software

• The Cline plugin in VS Code (the Roo Code plugin can also achieve similar functionality)

• The large model Deepseek R1 (just to test the capabilities of the latest R1) and its API (if you know magic, I also recommend Google’s Gemini 2.0 API)

Below, I will take VS + Cline + Deepseek as an example to write a small image compression tool and teach you how to build a Cursor-like software on your own.

1. Download Visual Studio Code Software

1. Visit code.visualstudio.com/download and select the Windows version for download.

2. After downloading, install it normally.

2. Install the Cline Plugin in VS Code

1. Open the VS Code software,find the Extensions icon in the left sidebar, select it, and enter “Cline” in the search box. Select the Cline plugin (with the robot head logo) and click install.

The second plugin in the image, Roo Code, is also a plugin with functionality very similar to Cline, and installing it can achieve similar results, serving as an alternative.

2. After successful installation, a robot head logo will appear in the left sidebar,click on it and then click the gear icon in the upper right corner to configure.

3. The most important configuration is to select the API, where we have two options:The first is to select Deepseek, which requires the Deepseek API(how to obtain the API will be covered in the following steps). The currently available large models are Deepseek-Chat (which should be V3) and Deepseek-Reasoner (which should be the recently popular R1). The reason for choosing Deepseek is not only because of its strong model capabilities but also its affordability (specific pricing will be explained in later steps).

The second option is to select Google Gemini, which requires the Google Gemini API (how to obtain the API will be covered in the following steps). There are many large models currently available, and it is recommended to choose gemini-2.0-flash-thinking-exp-1219, which has a high response speed and good output results. The main reason for choosing Google Gemini is that it is free, but the downside is that you need to know some magic to obtain the API and use it.

4. If you want to interact with Cline in Chinese during subsequent programming, please enter “Please respond in Chinese” in Settings – Personalization.

3. Obtain the Large Model API

Obtaining the Deepseek API

1. Visit the Deepseek open platform at platform.deepseek.com, register and log in. New users receive a 10 RMB gift upon registration. You can see that yesterday afternoon I used ds-chat (which is V3) for programming and consumed 35,000 tokens, resulting in a cost of 3 cents. Therefore, it is justified to call Deepseek the “Pinduoduo of the large model field”.

2. Click on API Keys in the left sidebar, click Create API Key, name the API, and click the Create button.

The API KEY has been successfully created. Please note,it is essential to copy the API key and paste it in a safe place. If you do not save it at this step, you will not be able to retrieve it later.

Obtaining the Google Gemini API

1. Use magic to visit aistudio.google.com, click the Get API Key button on the left. Click the Create API Key button.

In the pop-up dialog, select a project and click Create to generate the API. There is also a copy button for one-click copying.

The advantage of Google is that if you do not save upon creation, you can still retrieve it later by clicking to query.

4. Set the API in Cline

Return to the VS Code software and fill in the ds or Google API we obtained in the Cline plugin settings.

Comparing the pricing of Deepseek, the latest R1 has a maximum output of 8000 tokens, with cache write costing $0.55 per million tokens (4 RMB), cache read costing $0.14 per million tokens (1 RMB), and output costing $2.19 per million tokens (14 RMB).

If using Deepseek V3, cache write costs $0.14 per million tokens (1 RMB), cache read costs $0.01 per million tokens (7 cents), and output costs $0.28 per million tokens (2 RMB).

Of course, at this stage, Google is still the most cost-effective option, as it currently offers a free plan and has not yet cost anything.

5. Write an Image Compression Tool Website Using the Combination of VS + Cline + Deepseek

1. Create a folder on your local computer named “tupianyasuo”

2. Open it in VS (menu File – Open File or shortcut Ctrl + O)

3. To facilitate communication with Cline later, you can check the auto-approve option for Cline, so you won’t need to manually confirm each step.

The Max requests default setting is 20, which means that Cline needs to get manual authorization every 20 API calls (by clicking the approve button). If you find this cumbersome, you can set this value higher.

4. Similar to Cursor’s chat and composer, Cline is also divided into plan and act. In simple terms, plan is to first discuss requirements with AI, allowing it to help plan how to implement these requirements, including functional requirements, technology selection, general steps, and ideas. Once you have discussed enough, you can switch to act to complete the specific code. Recently, due to the popularity of Deepseek, the server resources may be insufficient, and the API requests keep spinning, so let’s try Google instead.

The key in the plan is to leverage AI thinking to refine your ideas, which is also why I’ve recently enjoyed using DS R1. Observing how AI articulates its thought processes is actually very beneficial for my own thinking. This reflects the evolution of using AI from “AI learning human thinking” to “humans learning AI thinking”. Just like AlphaGo, which emerged and later human experts began to learn from AI’s strategies to improve their own chess skills.

We can see that Cline has created three files (HTML, CSS, JavaScript) to implement the website functionality.

Once completed, Cline informs that testing can proceed.

5. Test the practical effect by uploading an image. It successfully completed compression, but the function to adjust the compression ratio does not work. Don’t worry, this is also a characteristic of AI programming; sometimes it cannot be achieved in one go and requires multiple adjustments.

I previously used Cursor’s Claude large model to create a similar website, which was successful on the first attempt, with both the compression ratio and adjustment functions performing better than this trial. I initially believe that the difference may be due to the capabilities of the large models; it is said that Claude may currently be the best in programming implementation.

Conclusion:

• The combination of VS Code + Cline + Deepseek/Google Gemini can achieve Cursor-like AI programming functionality;

• The APIs for Deepseek and Gemini need to be obtained separately; Deepseek V3 is very affordable, while R1 is slightly more expensive; Gemini requires magic and is currently free, but there may be prompts indicating quota exhaustion during use.

• This alternative combination does not perform as well as the previous Cursor + Claude, which may be related to the capabilities of the large models; overall, there is still some gap compared to Cursor.

• The learning threshold for AI programming is not high; experience needs to be accumulated through continuous practice to improve one’s ability to collaborate with AI.