把桌面 AI 软件变成本地 API：手把手反代教程

① 桌面版登录  →  凭证（token）写入本地文件
② 反代启动    →  读取凭证文件，拿到 token
③ 下游请求    →  反代补上协议要求的请求头 → 转发给上游 → 原样返回结果

python -c "import json,io;d=json.load(io.open(r'<安装目录>\resources\app.asar.unpacked\cli\product.json',encoding='utf-8'));print('endpoint =',d.get('endpoint'));print('productName =',d.get('productName'))"

import base64, json

tok = "<你的 accessToken>"
p = tok.split('.')[1]
p += '=' * (-len(p) % 4)
print(json.loads(base64.urlsafe_b64decode(p))['iss'])

Get-ChildItem "$env:LOCALAPPDATA" -Recurse -Filter "*.info" -ErrorAction SilentlyContinue |
  Where-Object { $_.FullName -match "auth" } |
  Select-Object -ExpandProperty FullName

{
  "auth": {
    "accessToken": "eyJhbGciOi...",
    "refreshToken": "...",
    "expiresAt": 1234567890
  },
  "account": {
    "uid": "...",
    "email": "..."
  }
}

# 列出代码里所有外部地址
grep -ohE "https?://[a-zA-Z0-9._/-]+" *.py | sed 's|\(https\?://[^/]*\).*|\1|' | sort -u

# 找危险调用
grep -nE "\beval\(|\bexec\(|pickle|subprocess|os\.system|upload|webhook" *.py

# 装依赖（一般就这两个）
pip install fastapi uvicorn

# 启动（用 -u，否则日志被缓冲，看不到实时输出）
python -u server.py

Uvicorn running on http://127.0.0.1:8787

import <反代项目的模块名> as _cda

# (a) 强制正确的上游 host
_orig_init = _cda.ApiClient.__init__
def _patched_init(self, base_url, token_info, safe_mode=False):
    _orig_init(self, "https://www.<你的网关域名>", token_info, safe_mode)
_cda.ApiClient.__init__ = _patched_init

# (b) 补上 User-Agent
_orig_headers = _cda.ApiClient._build_headers
def _patched_headers(self, extra=None, stream=False, model=""):
    h = _orig_headers(self, extra=extra, stream=stream, model=model)
    h.setdefault("User-Agent", "<产品标识>/0.0.0")
    return h
_cda.ApiClient._build_headers = _patched_headers

# (c) 保证首条是 system prompt，并把模型名映射成档位名
_orig_chat = _cda.ApiClient.chat_completion
def _patched_chat(self, messages, model="", **kw):
    model = {"deepseek-v4-flash": "fast-model", "deepseek-v4-pro": "primary-model"}.get(model, model)
    if not messages or (messages[0] or {}).get("role") != "system":
        messages = [{"role": "system", "content": "You are a helpful assistant."}] + list(messages)
    return _orig_chat(self, messages, model=model, **kw)
_cda.ApiClient.chat_completion = _patched_chat

# (d) 只暴露账号真实可用的档位名
TIERS = {"fast-model", "balanced-model", "primary-model", "deep-model"}
_cda.KNOWN_CHAT_MODELS = set(TIERS)
_cda.ALL_SUPPORTED_MODELS = set(TIERS)
_cda.FREE_MODELS = set()

400 model [xxx] service info not found
Currently supported models for your account:
  - fast-model
  - balanced-model
  - primary-model
  - deep-model

curl -s --noproxy '*' http://127.0.0.1:8787/health

{"status":"ok","<产品前缀>_configured":true}

curl -s --noproxy '*' http://127.0.0.1:8787/v1/models \
  -H "Authorization: Bearer <本地key>"

curl -s --noproxy '*' -X POST http://127.0.0.1:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <本地key>" \
  -d '{"model":"fast-model","messages":[{"role":"user","content":"1+1=?"}],"stream":false}'

# 找到旧实例
netstat -ano | grep :8787

# 杀掉（把 <pid> 换成上面查到的进程号）
taskkill /PID <pid> /F

# 按原来的方式重新启动
python -u server.py

import json, io, glob, os, base64

d = os.path.expandvars(r'%LOCALAPPDATA%\<产品目录>\Data\Public\auth')
for f in glob.glob(d + r'\*.info'):
    j = json.load(io.open(f, encoding='utf-8'))
    tok = (j.get('auth') or {}).get('accessToken', '')
    p = tok.split('.')[1]
    p += '=' * (-len(p) % 4)
    pl = json.loads(base64.urlsafe_b64decode(p))
    print(f)
    print('  iss     =', pl.get('iss'))
    print('  account =', j.get('account'))

# 添加转发规则：<宿主机IP>:8787 → 127.0.0.1:8787
netsh interface portproxy add v4tov4 listenaddress=<宿主机IP> listenport=8787 connectaddress=127.0.0.1 connectport=8787

# 放行防火墙
netsh advfirewall firewall add rule name="AI Proxy 8787" dir=in action=allow protocol=TCP localport=8787

# 查看已生效的规则
netsh interface portproxy show all

curl -s http://<宿主机IP>:8787/health

帮我把这台电脑上已登录的 <软件名> 桌面版做成一个本地 OpenAI 兼容接口，你全程执行。

1. 找到它的安装目录，读 product.json，拿到 endpoint 和产品标识
2. 在 %LOCALAPPDATA% 下搜路径含 auth 的 *.info 文件，读出 accessToken
3. 从 GitHub 找 <软件名>2api 项目并下载，先做安全检查（列出代码里所有外部域名，应该只有 localhost 和官方域名）
4. 装依赖，写一个启动器 monkey-patch 四处不匹配：上游 host、User-Agent、首条 system prompt、模型名映射
5. 启动服务，三步验证：/health、/v1/models（应只有 4 个档位名）、真实调用一次
6. 把 base_url、本地 key 和可用模型名给我

报错按 host → User-Agent → 请求格式 → 模型名 的顺序排查，需要我做什么直接说。