🧠 Cursor DeepSeek V4 Proxy

One-click fix for the reasoning_content error when using DeepSeek V4 in Cursor, say goodbye to Rate limit exceeded, and make AI Agent mode run stably.

📌 What problems can this project solve for you?

If you frequently encounter any of the following errors when calling DeepSeek V4 (Pro / Flash) in Cursor:

Provider returned error: The reasoning_content in the thinking mode must be passed back to the API.
User API Key Rate limit exceeded (error despite plenty of quota left)
AI Model Not Found: deepseek-v4-pro (background task reports invalid model name)
The first round of chat works fine, but errors and interruptions start from the second round

No need to struggle anymore, follow this guide and solve it completely in 5 minutes.

⚠️ Fixed on 2026-05-14: Pure text models now automatically filter images, resolving 502 errors. See Changelog for details.

✨ Core Features

✅ Automatic caching & returning of reasoning chains: no more errors caused by missing reasoning_content.
✅ Intelligent rate limiting: built-in token bucket to prevent burst concurrent requests from maxing out free quota.
✅ Supports streaming output: does not affect Cursor’s typewriter rendering effect.
✅ One-click start script: works on Windows / macOS / Linux, just double-click to run.
✅ Transparent logs: terminal shows request status in real-time for easy troubleshooting.
✅ Zero intrusion: no need to modify Cursor program files, only change one Base URL.

---

🖥️ Supported Environments

Only requirement: Install Python 3.8 or higher (be sure to check Add Python to PATH during installation).

🚀 Super Detailed Three-Step Getting Started (For Beginners)

Step 1: Download the project and install dependencies

Download the ZIP package of this project repository and extract it locally (please do not include Chinese characters in the path).
Enter the extracted folder, type cmd in the folder address bar and press Enter to open the command prompt.
Run the following command to install dependencies:

bash
    pip install -r requirements.txt


    If you see the prompt

pip is not an internal command, please reinstall Python and check Add to PATH.

`Step 2: Start Local Proxy + Tunnel`

You need a tunnel to generate a public address (Cursor restricts access to localhost).

#### 🪟 Windows Users

Make sure the folder contains cloudflared-windows-amd64.exe (if not, please download it from the Cloudflare official site).

Double-click to run start_proxy.bat.

Two windows will pop up. In the tunnel window, find a URL like https://xxx.trycloudflare.com and copy it.

#### 🍎 macOS / Linux Users
 Open the terminal, navigate to the project directory, and run:

bash
    bash start_proxy.sh

Wait a moment, then copy the terminal output URL https://xxx.trycloudflare.com.

> ⚠️ Note: Do not close the window. The tunnel address changes each time it restarts, but it remains valid as long as the window stays open.
Step 3: Configure Cursor

Open Cursor settings: press Ctrl+Shift+P → type Cursor Settings.


 Go to the Models tab.

In "Override OpenAI Base URL", paste the copied address, and add /v1 at the end:

For example: https://xxxxxx.trycloudflare.com/v1


 Enter your DeepSeek API Key in the API Key field.
 Completely exit and restart Cursor.

---
🛡️ Fix "Model name not valid" Error
If you encounter an error when running Apply or background tasks, follow these steps:

Press Ctrl+Shift+P, type Preferences: Open User Settings (JSON) and press Enter.

Add the following configuration inside the JSON braces {}:

json
    "cursor.models": {
        "deepseek-v4-pro": {
            "provider": "openai",
            "apiBase": "https://你的隧道地址.trycloudflare.com/v1",
            "apiKey": "你的DeepSeek API Key"
        }
    }


 Save and restart Cursor.
---
❓ FAQ

🔁 What if the tunnel address changes?
A new address is generated every time the script restarts. You just need to retrieve it again and update the Base URL in Cursor.


💸 Still getting "Rate limit exceeded"?
DeepSeek's free tier has a very low frequency limit. You can edit

proxy.py and reduce the 5 in TokenBucket(rate=5/60.0, capacity=5) (e.g., to 3

) to forcibly lower the request frequency.


🚫 Must I use a tunnel? Can I connect to localhost?
Cursor forbids direct connections to

localhost

 for security reasons. Cloudflare Tunnel is currently the simplest, free, and registration-free penetration solution.


🧪 Does the proxy affect the model's intelligence?
In 99% of scenarios, it is imperceptible. The proxy only automatically completes when the model "forgets" to return the thought chain, ensuring the conversation is uninterrupted.


🧪 Why is there only 200k context, not 1M?
This is the default limitation of Cursor, not the proxy issue.

Cursor uses a 200k context window by default. To enable the 1M context supported by the model, you need to turn onMax Modein Cursor Chat. Path: Chat -> Model Selector -> Max Mode

`⚙️ Advanced Customization`

Change Upstream: Modify the UPSTREAM_URL in proxy.py.


Fixed Domain: If you have your own domain, you can configure a Cloudflare Named Tunnel (Persistent Tunnel).

  
🆕 Changelog  
2026-05-14: Plain Text Model Image Compatibility Fix

Fix Details When clients like Cursor send messages containing images todeepseek-v4-pro (plain text model), the proxy layer now automatically filters out the image_urlcontent blocks, keeping only the text part. If a message consists entirely of images (no text), it is replaced with a prompt text to avoid the model receiving an empty message. This fix resolves the 502 error caused by DeepSeek API returning"unknown variant image_url, expected text".

Additional Improvements

Unified upstream error response format, so clients receive structured JSON errors (instead of raw error messages or blank pages).

Image compatibility logic does not affect inference cache (reasoning_content), and historical message handling remains normal.

------  
Acknowledgments  
Thanks to @BG-ah for reporting the rate limit issue in Issue #4, and to @CH-nolyn for participating in the discussion. Your feedback directly drove this compatibility fix, making the proxy run more stably under the plain text model.  
------

💡 Tip: If you want the model to truly understand image content, please change the model field in your request to the multimodal-supported deepseek-chat`, and ensure your DeepSeek account has the corresponding permissions enabled.

--- Tranlated By Open Ai Tx | Last indexed: 2026-06-24 ---

cursor-deepseek-v4-proxy

🧠 Cursor DeepSeek V4 Proxy

📌 What problems can this project solve for you?

✨ Core Features

🖥️ Supported Environments

🚀 Super Detailed Three-Step Getting Started (For Beginners)

Step 1: Download the project and install dependencies

Step 2: Start Local Proxy + Tunnel

Step 3: Configure Cursor

🛡️ Fix "Model name not valid" Error

❓ FAQ

⚙️ Advanced Customization

🆕 Changelog

2026-05-14: Plain Text Model Image Compatibility Fix

Acknowledgments

`Step 2: Start Local Proxy + Tunnel`

`⚙️ Advanced Customization`