Skip to content
BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
client
client
 
 
server
server
 
 
.gitignore
.gitignore
 
 
README.md
README.md
 
 
preview.png
preview.png
 
 
start-dev.ps1
start-dev.ps1
 
 

Repository files navigation

🧪 AI API Tester

一个本地跑的 AI API 调试工具。有界面，填字段比写 JSON 省事，切换模型/渠道点一下就切；请求体和返回体完全透明 —— 你发出去什么、收回来什么，全部摊开，方便调试。

A local AI API debugging tool. With a UI — filling fields beats hand-writing JSON, switching models/providers is one click; full request/response transparency — everything you send and receive is fully laid out for debugging.

支持 OpenAI Chat Completions、OpenAI Responses、Anthropic Messages 三种协议，以及兼容这三家的中转。

Supports OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and any compatible relay.

Preview

⚠️ 这个 README 是 AI 写的，可能有幻觉。具体行为请以代码为准，描述跟实际对不上的地方欢迎提 issue / 直接改。

虽然俺的本职工作是UI和用户体验设计, 但确实在这里比较偷懒, 只和Agent打磨了思路, UI和细节都没有精力打磨, 凑合用吧

My day job is UI/UX design, but I got lazy here — just brainstormed the ideas with an AI agent, didn't put any effort into polishing the UI or details. Use it as-is.

This README is AI-generated and may hallucinate. Code is the source of truth — open an issue or fix directly when descriptions diverge from reality.

💡 为什么做这个

从真实痛点里长出来的。

最早只是想给 opencode 里的模型开 thinking 模式，折腾半天发现：同样是 OpenAI Chat Completions 这个"通用协议"，各家字段还是不一样 —— OpenAI o-series 要 reasoning_effort，Anthropic 要 thinking.type + budget_tokens，字节豆包、智谱 GLM、月之暗面 MiniMax 又各自有别的字段名。返回也乱：推理数据有的包在 response 标签里，有的在 reasoning_content 字段，有的只在 SSE 流的某个 event 里闪过。想调试的话 —— opencode、Cline 这些上层工具不给你看真实的 HTTP 请求体和返回数据，debug 全靠猜。

所以做了这个过程透明的工具：你发出去的 URL / headers / body，和收回来的 status / headers / body / 原始 SSE 帧，全部摊在你面前。字段怎么配，自己试，立刻看到结果。

Born from real pain. I just wanted to enable thinking mode for a model inside opencode, and discovered that even within the supposedly "universal" OpenAI Chat Completions protocol, every vendor has different fields — OpenAI o-series wants reasoning_effort, Anthropic wants thinking.type + budget_tokens, Doubao/GLM/MiniMax each invent their own. Responses are equally messy: reasoning data hides inside response tags, or in a reasoning_content field, or flickers past in one specific SSE event. Want to debug? Upper-layer tools like opencode and Cline don't show you the real HTTP request/response — it's all guesswork.

🚀 启动

这个项目天生适合让 AI agent 帮忙跑 —— 启动、清理端口、健康检查这些重复劳动全部可以交给 agent。

需要 Node.js 18+。

1. 📥 下载到本地

git clone <repo-url>

clone 下来后什么都不用手动改，把项目目录丢给 AI agent 就行。

2. 🤖 让 AI agent 帮你启动

把项目路径告诉 agent（OpenCode、Claude Code、Cursor 等），然后对它说一句：

帮我启动这个项目

agent 会自动：

  1. 装 server/ 和 client/ 的依赖
  2. 清理 3001 / 28001 端口的旧进程
  3. 启动后端（Express，端口 3001）
  4. 启动前端（Vite，端口 28001）
  5. 健康检查两个服务
  6. 报告本机 + 局域网访问地址

3. 🌍 打开浏览器

agent 启动完会告诉你访问地址，默认是 **http://localhost:28001/**。

🛟 兜底：手动启动

如果手头没 AI agent：

# 终端 1
cd server && npm install && npm run dev

# 终端 2
cd client && npm install && npm run dev

Windows 用户可以直接双击项目根的 start-dev.ps1 一键启动。

This project is built to be launched by an AI agent — starting servers, killing port conflicts, and health-checking can all be delegated. Requires Node.js 18+. After cloning, hand the directory to your agent and say "Start this project for me". The agent will: install deps, kill stale processes on ports 3001/28001, start backend (Express, port 3001) and frontend (Vite, port 28001), health-check both, and report local + LAN URLs. Manual fallback: run npm install && npm run dev in server/ and client/ separately, or double-click start-dev.ps1 on Windows.

🖥️ 界面

三栏布局，分隔线可拖：

┌─────────────┬───────────────────┬───────────────────┐
│  Providers  │     Request       │     Response      │
│  渠道列表    │   构造请求         │   看响应          │
└─────────────┴───────────────────┴───────────────────┘

👈 左 — Providers —— 渠道列表（名字 / 协议 / baseUrl / API key / 模型列表），点一条加载到中间栏。

✏️ 中 — Request —— 协议、模型、baseUrl、采样参数、API key、reasoning 模板、messages、system、extra body。

👀 右 — Response —— 流式时实时显示 Output / Reasoning；完成后展开看 Sent Request、Received Response、原始 SSE Frames。

Three-column layout with draggable dividers. Left — Providers: provider list (name / protocol / baseUrl / API key / models), click to load. Middle — Request: protocol, model, baseUrl, sampling, API key, reasoning template, messages, system, extra body. Right — Response: real-time Output/Reasoning while streaming; expand panels to inspect Sent Request, Received Response, and raw SSE Frames.

🔌 三个协议

协议 路径
openai-completions {baseUrl}/v1/chat/completions
openai-responses {baseUrl}/v1/responses
anthropic {baseUrl}/v1/messages

Three protocols supported: OpenAI Chat Completions (/v1/chat/completions), OpenAI Responses (/v1/responses), Anthropic Messages (/v1/messages).

🧠 Reasoning 模板

不同家的"先想后答"字段不一样，工具用模板抹平。预置了四个模板：

模板 字段
OpenAI Reasoning Effort reasoning_effort（none / minimal / low / medium / high / xhigh）
Claude Effort thinking.type（adaptive）+ thinking.display（summarized）+ output_config.effort（low / medium / high / xhigh / max）
火山方舟 Coding Plan thinking.type（enabled / disabled）+ reasoning_effort（minimal / low / medium / high / max）
MiniMax thinking.type（adaptive / disabled）+ reasoning_split（true / false）

下拉框选一个就行，自定义后可以保存为新模板。

Every vendor names their "think first" fields differently — templates paper over the differences. Four are bundled: OpenAI Reasoning Effort (reasoning_effort), Claude Effort (thinking.type + thinking.display + output_config.effort), Volcengine Ark Coding Plan (thinking.type + reasoning_effort), MiniMax (thinking.type + reasoning_split). Pick from the dropdown, or save your own.

💾 数据存在哪

数据 位置
Provider 配置（含 API key） server/data/configs.json
Reasoning 模板 server/data/reasoning-templates.json
请求参数缓存 浏览器 localStorage

server/data/ 已经在 .gitignore。

Provider configs (with API keys) → server/data/configs.json, reasoning templates → server/data/reasoning-templates.json, per-model request cache → browser localStorage. server/data/ is already in .gitignore.

🔒 安全提醒

  • 🚫 API key 明文存在本地 JSON。别 commit、别放公网。
  • 🚫 工具没有任何鉴权，只能本机用。
  • 🔄 不小心推了 key → 立刻去 provider 后台 rotate。

API keys are stored as plaintext JSON locally — never commit them, never expose on the public internet. No authentication — local use only. Accidentally pushed a key? Rotate it immediately on the provider dashboard.

❓ 常见问题

Q: 启动后浏览器打不开？ A: 看终端有没有 EADDRINUSE，端口被占了。换端口或者 kill 占用进程。

Q: 提示 "API key is required"? A: 中间栏 🔑 API Key 是空的，或者你用的是占位 provider。换成你自己的 key。

Q: 提示 "Model is required"? A: Endpoint 区域的 Model 字段是空的。下拉框选一个，或者自己输入模型名。

Q: 中转站拉不到模型列表？ A: 不是所有中转都开放 /v1/models。手动填模型名就行。

Q: 报 "Invalid baseUrl"? A: baseUrl 必须带协议头，比如 https://api.openai.com，不能只填 api.openai.com。

Q: 能同时给同事用吗？ A: 不能。它没有任何鉴权，只适合本机调试。要团队用，得自己在外面套个登录层。

Browser won't open? → Check for EADDRINUSE, the port is taken. "API key is required"? → Fill in your key. "Model is required"? → Pick a model. Relay can't return model list? → Type manually. "Invalid baseUrl"? → Add https:// scheme. Share with teammates? → No, no auth — local only.

📁 项目结构

apitest/
├── client/                      # 前端 (Vite + React)
│   └── src/
│       ├── App.tsx
│       ├── components/
│       │   ├── ConfigList.tsx   # 左栏 provider CRUD
│       │   └── Tester.tsx       # 中右栏：请求编辑器 + 响应显示
│       └── hooks/useColumnResizer.ts
├── server/                      # 后端 (Express)
│   ├── src/index.ts             # API 路由 + SSE 解析 + 协议转换
│   └── data/                    # gitignored
└── start-dev.ps1

🛠️ 技术栈: React 18 + Vite 5 + TypeScript 前端，Express 4 + tsx 后端，本地 JSON 存储，SSE 流式。

Stack: React 18 + Vite 5 + TypeScript (frontend), Express 4 + tsx (backend), local JSON storage, SSE streaming.

🌐 API 端点

方法 路径 说明
POST /api/proxy 转发请求到上游 LLM（支持流式）
GET /api/models?baseUrl=...&apiKey=... 拉模型列表
GET/POST/PUT/DELETE /api/configs[/:id] Provider CRUD
PATCH /api/configs/:id/endpoint 部分更新 baseUrl / basePath
GET/POST/PUT/DELETE /api/reasoning-templates[/:id] Reasoning 模板 CRUD

POST /api/proxy — proxy request to upstream LLM (streaming supported). GET /api/models — fetch model list. GET/POST/PUT/DELETE /api/configs[/:id] — provider CRUD. PATCH /api/configs/:id/endpoint — partial baseUrl/basePath update. GET/POST/PUT/DELETE /api/reasoning-templates[/:id] — reasoning template CRUD.

📦 POST /api/proxy 请求体：

{
  "protocol": "openai-completions",
  "stream": false,
  "model": "...",
  "messages": [{ "role": "user", "content": "..." }],
  "system": "optional system prompt",
  "maxTokens": 1024,
  "temperature": 1,
  "apiKey": "sk-...",
  "baseUrl": "https://api.example.com",
  "extraBody": {},
  "reasoningFields": [
    { "name": "effort", "value": "high", "target": "reasoning_effort" }
  ]
}

📜 License

暂未指定。如需对外发布，请补充 LICENSE 文件后再行声明。

Not specified yet. Add a LICENSE file before public release.

About

A local AI API debugging console — Postman for LLM APIs. Unified OpenAI Chat, OpenAI Responses, and Anthropic Messages, with pluggable baseUrl, live streaming, and reasoning field templates.

Resources

Readme
Activity

Stars

2 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages