GptBypass

A local proxy for OpenAI-compatible API that detects model refusals and automatically rewrites prompts for retry.

中文说明

---

Features

• Transparent proxy — Forwards all /v1/* requests to your target model endpoint
• Fake /v1/models — Returns a local model list so clients can check connectivity
• Refusal detection — Scans the model's last reply against a configurable denylist
• Automatic prompt rewriting — When refusal is detected, an auxiliary model rewrites the last user message and retries
• GUI & CLI — Comes with both a desktop GUI (CustomTkinter) and a CLI mode
• Rust-accelerated filtering — Optional Aho-Corasick matcher via PyO3 for high-performance keyword detection
• Streaming support — Fully handles SSE streaming responses with buffered analysis

How It Works

Client  →  GptBypass Proxy  →  Target Model
                ↓                     ↓
         Denylist Check      Model Response
                ↓                     ↓
         Hit? → Rewrite      Clean → Forward
         Prompt via Aux       to Client
         Model → Retry

1. Client sends a request through the proxy
2. First attempt goes straight to the target model (no rewrite)
3. Proxy buffers and analyzes the response
4. If the response matches any denial keyword → discard, rewrite the prompt via the auxiliary model, and retry
5. If the response is clean → forward to client

Quick Start

Prerequisites

• Python 3.10+
• A target model API endpoint (OpenAI, local, or any compatible API)
• An auxiliary model API endpoint for prompt rewriting

Installation

git clone https://github.com/YOUR_USERNAME/GptBypass.git
cd GptBypass
pip install -r requirements.txt

Configuration

Copy the example config and fill in your API details:

cp config.example.json config.json

Edit config.json:

{
    "target_model": {
        "model": "gpt-5.4",
        "message_type": "responses",
        "reasoning_depth": "high",
        "baseurl": "http://your-target-api/v1",
        "apikey": "your-target-api-key"
    },
    "optimization_model": {
        "model": "your-aux-model",
        "baseurl": "https://openrouter.ai/api/v1",
        "apikey": "your-aux-api-key",
        "max_retries": 5
    },
    "response_filter": {
        "denylist": ["抱歉", "不能协助", "I'm sorry", "I cannot assist"]
    }
}

⚠️ Never commit your config.json with real API keys. It is already in .gitignore.

Run

GUI mode:

python gui_app.py

CLI mode:

# Start proxy server
python cli_app.py serve --host 127.0.0.1 --port 8999

# Or use the simple runner
python run_proxy.py --host 127.0.0.1 --port 8999

Then point your client (Cursor, Continue, etc.) to http://127.0.0.1:8999/v1.

Configuration Reference

target_model

The model you ultimately want to use.

Field	Description
model	Model name forwarded to the target backend
message_type	API type: responses or chat.completions
reasoning_depth	Reasoning effort level (e.g. high, medium, low)
baseurl	Target API endpoint URL (include /v1)
apikey	Target API key (leave empty if not needed)

optimization_model

The auxiliary model used to rewrite prompts when refusal is detected.

Field	Description
model	Auxiliary model name
baseurl	Auxiliary API endpoint
apikey	Auxiliary API key
system_prompt	System prompt for the rewriting model
log_full_refined_content	Log full rewritten prompts (true/false)
only_main_user_request	Only rewrite main requests, skip auxiliary ones like title generation
max_retries	Maximum retry attempts after refusal

response_filter

Field	Description
denylist	List of keywords/phrases. If the model's last reply contains any, it triggers a rewrite + retry

Rust Filter (Optional)

For better keyword matching performance, build the native Rust module:

cd rust_filter
pip install maturin
maturin develop --release

This provides Aho-Corasick based matching via PyO3. The proxy automatically falls back to Python if the Rust module is unavailable.

Building Standalone Executables

CLI executable:

.\build_exe.ps1

GUI executable:

.\build_gui_exe.ps1

Outputs are placed in dist/ and dist_gui/ respectively.

Project Structure

GptBypass/
├── proxy/
│   ├── __init__.py
│   └── main.py              # Core proxy engine (FastAPI)
├── rust_filter/
│   ├── src/lib.rs            # Rust Aho-Corasick filter (PyO3)
│   └── Cargo.toml
├── gui_app.py                # Desktop GUI (CustomTkinter)
├── cli_app.py                # CLI entry point
├── run_proxy.py              # Simple proxy runner
├── app_defaults.py           # Default configuration
├── config.example.json       # Example config (no secrets)
├── guide.html                # Usage guide (Chinese)
├── app_icon.ico              # Application icon
├── ico.png                   # Icon PNG
├── tests/
│   ├── test_app.py           # Unit & integration tests
│   └── test_client.py        # Simple client test
├── build_exe.ps1             # Build CLI exe
├── build_gui_exe.ps1         # Build GUI exe
├── requirements.txt
├── LICENSE
└── README.md

API

The proxy exposes these HTTP endpoints:

Method	Path	Description
GET/HEAD	/v1/models	Returns a fake model list for client compatibility
*	/{path:path}	Proxies everything else to the target model

Request Classification Headers

You can control bypass behavior per-request using custom headers:

Header	Value	Effect
x-jmp-main-request	true	Force treat as main request (always rewrite on refusal)
x-jmp-aux-request	true	Force treat as auxiliary request (skip rewrite)
x-jmp-skip-optimize	true	Skip optimization entirely

Logs

Default log file: proxy.log (alongside the executable)

Common log patterns:

• 请求进入 — New request received
• 首轮直发目标模型 — First attempt, sent directly
• 命中过滤 | #1 — Response matched denylist
• 改写 | #2 — Starting rewrite retry
• 放行 | stream=True — Clean response forwarded to client

License

MIT

---

中文说明

一个基于 FastAPI 的 OpenAI 兼容 API 本地代理工具，主要用于自动检测模型拒答并改写提示词重试。

核心流程

1. 客户端请求发往本地代理
2. 首轮请求原样转发目标模型
3. 代理检测模型回复是否命中拒答关键词
4. 命中 → 丢弃该回复，调用辅助模型改写用户提示词后重试
5. 未命中 → 正常返回给客户端

快速开始

pip install -r requirements.txt
cp config.example.json config.json
# 编辑 config.json 填入你的 API 地址和密钥
python gui_app.py   # 图形界面
python cli_app.py serve  # 命令行

配置说明

• target_model — 目标模型接口配置（最终请求转发目标）
• optimization_model — 辅助改写模型配置（拒答后用于改写提示词）
• response_filter.denylist — 拒答关键词列表，命中任一关键词即触发改写重试

详见 guide.html 或复制 config.example.json 查看完整配置项。