Free Claude Code Setup với DeepSeek V4 hướng dẫn chi tiết

Integrating DeepSeek with Cursor for Affordable AI-Assisted Coding

The Cost Problem and DeepSeek as a Solution

DeepSeek is a newly released model with performance nearly matching GPT-5.5 or Opus, but at a fraction of the cost.

• The creator begins by addressing a common frustration among Cursor Pro users: the rapid consumption of credits. Even with a $20 credit top-up, a single prompt to generate a slide can exhaust the balance, forcing users to seek more efficient alternatives. He notes that not everyone can afford the $200 Cursor Max plan, which he himself finds expensive. The core problem is the high cost of premium models like Opus 4.7, which can cost tens of dollars per four-hour session. DeepSeek V4 Pro emerges as the answer, offering comparable performance—especially in coding—while drastically reducing expenses. With 1.6 trillion parameters and a 1-million-token context window, it can process entire books. It is open-source, but the creator clarifies that running it locally requires expensive hardware (H100 chips). Instead, users can access it via DeepSeek's API, Nvidia's free tier, or OpenRouter, all at much lower prices. He claims the same workflow can cost 50 times less than using Opus, making high-quality AI coding accessible to everyone.

Cost Comparison and Performance Benchmarks

In a single 4-hour work session, I could spend tens of dollars on Opus, even more, but when we use DeepSeek, the cost drops dramatically.

• The presenter provides concrete cost comparisons to illustrate the savings. For example, Opus 4.7 charges approximately $1.40 per million input tokens and a steep $25 per million output tokens. In contrast, DeepSeek V4 Pro costs only $0.036 per million input tokens and $0.08 per million output tokens—roughly 700 times cheaper on output. The even more economical DeepSeek V4 Flash is priced at $0.028 and $0.027, respectively. These figures are shown on OpenRouter's pricing page and DeepSeek's own pricing page. Additionally, DeepSeek's benchmarks are nearly on par with the top models: its coding performance is close to Opus 4.7 and GPT-5.5, as indicated by green bars on its official chart. The presenter emphasizes that despite being slightly behind Opus in overall scores, the cost advantage makes it an ideal choice for regular coding tasks. He also notes that for the first few months, DeepSeek is offering a 75% discount, further reducing costs, and encourages users to take advantage of this promotion before May 31.

Introducing the Free Code Library as a Universal API Wrapper

This library runs a web server on localhost:8082, wrapping APIs from providers like OpenRouter and Nvidia, bundling multiple AI models into a single endpoint.

• The tutorial shifts to explaining the Free Code library, an open-source project with over 20,000 stars on GitHub. This library acts as a proxy that aggregates multiple AI model providers—such as DeepSeek, OpenRouter, and Nvidia—into a single local server. By redirecting Cursor's communication to this local endpoint, users can seamlessly switch between models without changing their workflow. The library handles API key management and provides a unified interface. The presenter highlights that this eliminates the need to commit to a single provider, allowing users to choose the cheapest or most suitable model at any time. For instance, if a user runs out of credits on DeepSeek, they can instantly switch to free models available on OpenRouter, such as the new Tensen 23 Preview or other inexpensive options. This flexibility is key to maintaining productivity while controlling costs. The library also supports customized environment variables, enabling advanced users to configure multiple keys and model preferences.

Step-by-Step Setup: Downloading, Configuring, and Running the Proxy

Copy the example environment file, remove the extension, and rename it to just .env, then fill in your API keys for DeepSeek, OpenRouter, or Nvidia.

• The creator provides a detailed, beginner-friendly walkthrough. First, users must clone the Free Code repository from GitHub (or download it as a ZIP) and install Python if not already present. Within the project folder, there is an example environment file (.env.example) that must be copied and renamed to .env. Users then insert their API keys—either from DeepSeek (by creating an API key via the DeepSeek website after a $5 deposit) or from OpenRouter. The presenter demonstrates obtaining a DeepSeek API key: after logging in, navigate to "Access API Key," generate a new key, and copy it. Then, in the .env file, set the deepseek_api_key variable. For OpenRouter, a similar key is added. Next, users must set the default model by specifying the provider prefix and model name—for example, "deepseek/deepseek-v4-pro" using the naming convention from the library's documentation. After saving the file, they run two commands: first to install the UV package manager (if on Windows), then to start the server with "uv run unicorn" (or equivalent). Once the server is running on localhost:8082, it is ready to accept requests from Cursor.

Practical Demonstration: Generating a Landing Page with DeepSeek

Create for me a landing page about a free two-day AI agent course, covering everything from basics to advanced, scheduled for next Friday and Saturday.

• With the proxy running, the presenter switches to Cursor and selects the DeepSeek V4 Pro model (visible in the model list after configuration). He then issues a simple prompt asking DeepSeek to create a landing page for a hypothetical free AI agent course. Within moments, the model generates a fully functional HTML page complete with hero section, course outline, schedule, benefits, and FAQ. The presenter shows the output—a clean, responsive design that looks professional. He also demonstrates how token usage is displayed in real-time (input tokens around 800). To further test capabilities, he asks DeepSeek to add CSS animations to the existing page, and the model quickly appends smooth fade-in effects and other transitions, making the site more visually appealing. The entire process uses minimal credits due to DeepSeek's low cost. This practical example proves that even a single prompt can produce a complete, polished webpage, highlighting the model's coding proficiency and the value of the integration. The presenter also mentions that he has additional skill sets for creating marketing content, which he offers for download.

Configuring VS Code Extension and Final Summary

Here we have something called 'cod environment variable.' I will paste this in, then restart Cursor for the extension to pick up the local proxy.

• For users who prefer a graphical interface over the terminal, the presenter explains how to configure the Cursor extension in VS Code. After opening settings, navigate to "Cursor Environment Variables" and paste the necessary configuration that points Cursor to the local proxy at localhost:8082. After saving and restarting the IDE, the new model list appears in the extension, including DeepSeek V4 Pro. This allows users to switch models directly from the Cursor interface without touching the terminal again. The tutorial concludes with a recap of the entire process: clone the Free Code library, register for DeepSeek or OpenRouter, obtain API keys, edit the .env file to set the keys and default model, start the server, and optionally configure the extension for convenience. The presenter emphasizes that this method provides access to virtually any AI model at extremely low cost—sometimes even free via OpenRouter's free tier. He invites viewers to explore his skill packs and join his community for more resources. The demonstration ends with a final note on the potential of integrating cost-effective Chinese models like DeepSeek into daily coding workflows, making advanced AI assistance affordable for everyone.