SAP Extension Creator

Overview

Create Super Agent Party extensions—self-contained packages that extend the platform with custom chat UI and tools. Two modes are supported:

• Static extension: Pure HTML/CSS/JS frontend, served directly by SAP from the extension folder
• Node.js extension: Full-stack with Express backend, auto-managed by SAP (npm install + node index.js <port>)

Both modes support MCP tool registration (the register_node_extension_mcp protocol message works for ANY extension via WebSocket, despite the "node" in its name).

Quick Decision Tree

User wants to create an extension?
├─ Only needs UI (chat, display, simple interactions)? → Static Extension
└─ Needs backend logic (API calls, DB, file processing)? → Node.js Extension

Core Files Every Extension Needs

File	Required	Purpose
package.json	✅	Metadata, dependencies, window config
index.html	✅	Main UI (full HTML page, single-file app)
index.js	Node only	Node.js entry point
node_modules/	Node only	Auto-installed by SAP via npm install

Workflow

Step 1: Gather Requirements

Ask the user:

1. Extension name? (hyphen-case, e.g., my-weather-widget)
2. Description? (one sentence)
3. Static or Node.js? (Node.js only if backend logic/server-side code is needed)
4. For Node.js: what npm dependencies?
5. Should it register custom tools for the AI? (works in both static and Node.js modes via WebSocket MCP)
6. GitHub repository URL? (optional, for updates)
7. Transparent window? (frameless, always-on-top — for mini widgets like music controllers)
8. Default window size? (width/height in pixels)

Step 2: Scaffold the Extension

Use the templates in assets/ as starting points:

• Static: Copy assets/static-template/
• Node.js: Copy assets/node-template/

Create the extension directory under the workspace (user will later install it into SAP's extensions/ folder).

Step 3: Write package.json

See references/package-json-spec.md for the complete field reference. Minimum:

{
  "name": "my-extension",
  "version": "1.0.0",
  "description": "What it does",
  "author": "your-name",
  "repository": "https://github.com/user/repo",
  "backupRepository": "https://gitee.com/user/repo",
  "category": "Tools"
}

For Node.js extensions, also include:

{
  "main": "index.js",
  "nodePort": 0,
  "dependencies": { "express": "^5.1.0" }
}

For transparent/frameless widgets (e.g., mini music controllers, floating panels):

{
  "transparent": true,
  "width": 280,
  "height": 80
}

When transparent: true, SAP creates a frameless, transparent, always-on-top window (see main.js open-extension-window handler). Use this for compact overlay widgets.

Step 4: Write index.html

The HTML page is rendered inside an Electron BrowserWindow (either directly or via an iframe). Key patterns:

• Self-contained: The extension is a single HTML file with all CSS/JS inlined or loaded from CDN. For Node.js extensions, static assets are served from the extension directory.
• Font Awesome: Use CDN to ensure reliable loading in both static and Node.js modes:

  <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css">
  

  Avoid relative paths like ../../fontawesome/ — these may work for static extensions but break for Node.js extensions (different serving paths).
• Dark/Light mode: Always support both (see "Theme & i18n" section below).
• i18n (Chinese/English): Always support bilingual UI (see "Theme & i18n" section below).
• WebSocket connection: Connect to ws://host/ws for messaging and MCP.
• Extension ID: Parse window.location.pathname for /extensions/{ext_id}/.
• Message rendering: Listen for messages_update and broadcast_messages events.
• Send user input: Send set_user_input then trigger_send_message.

Step 5: Write index.js (Node.js only)

See references/node-entry-spec.md for the full protocol. The entry point:

1. Receives a port number via process.argv[2]
2. Starts an Express server on that port at 127.0.0.1
3. Serves static files from its own directory
4. Exposes a /health endpoint for readiness checks
5. SAP reverse-proxies requests to the extension

Step 6: Implement Tool Registration (optional, works in both modes)

Extensions can register tools that the AI agent can call — via WebSocket in the frontend (both static and Node.js). The MCP lifecycle has three mandatory stages:

STARTUP  → ws.onopen         → registerMcpTools()
RUNTIME  → ws.onmessage      → handleMcpCall() when AI calls a tool
SHUTDOWN → window.beforeunload → unregisterMcpTools()

① Register on startup — always in ws.onopen, using a dedicated function:

function registerMcpTools() {
  getExtId();
  ws.send(JSON.stringify({
    type: 'register_node_extension_mcp',
    data: {
      ext_id: MY_EXT_ID,
      tools: [{
        name: `${MY_EXT_ID}_my_tool`,
        description: 'What this tool does (use the user\'s language)',
        parameters: {
          type: 'object',
          properties: {
            param1: { type: 'string', description: '...' }
          },
          required: ['param1']
        }
      }]
    }
  }));
}

② Handle tool calls — the AI agent calls your tool:

async function handleMcpCall(data) {
  const { ext_id, tool_name, tool_params, call_id } = data;
  if (ext_id !== MY_EXT_ID && !tool_name.includes(MY_EXT_ID)) return;
  // ... execute logic, then:
  ws.send(JSON.stringify({
    type: 'mcp_tool_result',
    data: { call_id, result: 'output' }
  }));
}

③ Unregister on shutdown — MUST send unregister_node_extension_mcp before the window closes:

function unregisterMcpTools() {
  if (ws && ws.readyState === WebSocket.OPEN) {
    ws.send(JSON.stringify({ type: 'unregister_node_extension_mcp', data: { ext_id: MY_EXT_ID } }));
  }
}
window.addEventListener('beforeunload', () => { unregisterMcpTools(); });

Key rule: Registration and unregistration MUST be in separate named functions (registerMcpTools / unregisterMcpTools), NOT inline code. This makes the lifecycle explicit and easy for AI to understand.

If an extension has no MCP tools, all three functions can be deleted.

See sap-lx-music/index.html for a complete real-world MCP implementation example (static extension with 12+ registered tools).

---

Theme & i18n (Dark/Light Mode + Bilingual)

Every extension should support dark/light mode and Chinese/English bilingual UI. Do NOT hardcode a single theme color scheme — use CSS variables so each extension can have its own identity.

CSS Variable Pattern

Define light theme in :root and override in body.dark:

:root {
  --bg: #ffffff;
  --bg-secondary: #f5f5f5;
  --text: #333333;
  --text-sub: #888888;
  --accent: #ec4141;        /* extension's own brand color */
  --accent-hover: #d73a3a;
  --border: rgba(0,0,0,0.08);
  --transition: 0.3s cubic-bezier(0.25, 0.1, 0.25, 1);
  --font: -apple-system, BlinkMacSystemFont, "SF Pro Display", "Helvetica Neue", sans-serif;
}

body.dark {
  --bg: #2b2b2b;
  --bg-secondary: #222222;
  --text: #e0e0e0;
  --text-sub: #888888;
  --border: rgba(255,255,255,0.06);
}

* { box-sizing: border-box; margin: 0; padding: 0; }
html, body {
  height: 100%; font-family: var(--font);
  background: var(--bg); color: var(--text);
  transition: background var(--transition);
}

Dark Mode Toggle

function initTheme() {
  const saved = localStorage.getItem('myext_dark');
  if (saved === 'dark' || (!saved && matchMedia('(prefers-color-scheme:dark)').matches)) {
    document.body.classList.add('dark');
  }
}

function toggleDarkMode() {
  const isDark = document.body.classList.toggle('dark');
  localStorage.setItem('myext_dark', isDark ? 'dark' : 'light');
}

i18n Pattern

const i18n = {
  zh: {
    welcome: '欢迎使用我的扩展',
    send: '发送',
    // ... all UI strings
  },
  en: {
    welcome: 'Welcome to My Extension',
    send: 'Send',
    // ...
  }
};

let lang = localStorage.getItem('myext_lang') || 'zh';
function t(k) { return i18n[lang]?.[k] || i18n.zh[k] || k; }

function toggleLanguage() {
  lang = lang === 'zh' ? 'en' : 'zh';
  localStorage.setItem('myext_lang', lang);
  updateAllTexts();  // re-render all i18n-dependent UI
}

When registering MCP tools, set description and parameters in the current user's language for better AI interaction.

---

Responsive Design

Every extension should work well across different window sizes. Critical patterns:

Viewport Meta (REQUIRED)

<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no" />

CSS Media Queries

Use breakpoints to adapt layout at small sizes:

@media (max-width: 900px) {
  /* stack layouts vertically, reduce padding */
}

@media (max-width: 600px) {
  /* hide secondary elements, compact controls */
}

Key responsive practices:

• Use vw units for widths as fallback (e.g., width: 65vw; max-width: 360px)
• Use flex layouts with flex-wrap that naturally adapt
• Hide non-essential elements on small screens (display: none)
• Reduce font sizes and padding at breakpoints

---

iframe Compatibility

Extensions may be rendered inside an iframe (depending on SAP's configuration). Ensure:

• Extension ID detection: Use window.location.pathname (works in both direct and iframe contexts):

  function getExtId() {
    try {
      const match = window.location.pathname.match(/\/extensions\/([^\/]+)/);
      return match ? match[1] : 'unknown';
    } catch(e) { return 'unknown'; }
  }
  

• WebSocket connection: Use location.host (not hardcoded):

  const proto = location.protocol === 'https:' ? 'wss:' : 'ws:';
  ws = new WebSocket(`${proto}//${location.host}/ws`);
  

• Window close: window.close() works in both direct and iframe contexts
• Avoid window.top / window.parent assumptions — your extension may be the top-level window
• Font Awesome via CDN ensures icons load regardless of serving path

---

Transparent Window / Compact Mode

When transparent: true is set in package.json, SAP creates a frameless transparent window. The extension must implement compact mode to work correctly.

How SAP Creates Transparent Windows

From main.js, when extension.transparent is true:

{
  frame: false,
  transparent: true,
  alwaysOnTop: true,
  skipTaskbar: false,
  hasShadow: false,
  backgroundColor: 'rgba(0, 0, 0, 0)',
}

Compact Mode CSS (REQUIRED for transparent extensions)

/* Transparent backgrounds */
body.compact { background: transparent !important; }
html.compact { background: transparent !important; }

/* Drag regions — make structural elements draggable for frameless windows */
body.compact header,
body.compact footer,
body.compact #inputBar {
  -webkit-app-region: drag;
}

/* Interactive elements MUST opt-out of drag */
body.compact button,
body.compact input,
body.compact textarea,
body.compact select,
body.compact a,
body.compact .compact-close-btn {
  -webkit-app-region: no-drag;
}

/* Compact close button (red circle, top-right) */
.compact-close-btn { display: none; }
body.compact .compact-close-btn {
  display: flex;
  position: absolute;
  top: 5px; right: 5px;
  width: 20px; height: 20px;
  background: rgb(255, 57, 57);
  border: none; border-radius: 50%;
  color: #fff;
  align-items: center; justify-content: center;
  font-size: 10px; cursor: pointer;
  transition: 0.2s;
  z-index: 100;
  -webkit-app-region: no-drag;
}
body.compact .compact-close-btn:hover { background: #ec4141; }

Compact Mode Detection (REQUIRED)

function checkCompactMode() {
  if (window.innerHeight < 200) {
    document.documentElement.classList.add('compact');
    document.body.classList.add('compact');
  } else {
    document.documentElement.classList.remove('compact');
    document.body.classList.remove('compact');
  }
}

function closeWindow() { window.close(); }

checkCompactMode();
window.addEventListener('resize', checkCompactMode);

Placing the Close Button

The close button HTML must be placed at the body level (not nested inside containers), typically right after <body>:

<body>
  <button class="compact-close-btn" onclick="closeWindow()" title="关闭窗口">
    <i class="fa-solid fa-xmark"></i>
  </button>
  <!-- rest of content -->
</body>

For transparent mini-widgets, you can also place the close button inside a content container and make it visible on hover — see sap-lx-music for this pattern.

---

Using iframes for Custom URL Schemes

If your extension needs to invoke custom protocol URLs (e.g., lxmusic://, myapp://), use a hidden iframe technique:

function invokeScheme(url) {
  let iframe = document.getElementById('scheme-invoker');
  if (!iframe) {
    iframe = document.createElement('iframe');
    iframe.id = 'scheme-invoker';
    iframe.style.display = 'none';
    document.body.appendChild(iframe);
  }
  iframe.src = url;
}

This avoids window.open() popup blockers and works reliably inside Electron.

---

WebSocket Protocol Reference

Message Type	Direction	Purpose
get_messages	→ SAP	Request current message history
messages_update	← SAP	Message list updated
broadcast_messages	← SAP	Broadcast message update
set_user_input	→ SAP	Update user input text
trigger_send_message	→ SAP	Send current input as user message
trigger_clear_message	→ SAP	Clear all messages
register_node_extension_mcp	→ SAP	Register MCP tools (works for static AND Node.js)
unregister_node_extension_mcp	→ SAP	Unregister on page close
mcp_registered	← SAP	Confirmation of registration
call_mcp_tool	← SAP	AI agent calls a registered tool
mcp_tool_result	→ SAP	Return tool execution result
trigger_close_extension	→ SAP	Request extension window close

---

Simple Chat HTTP API (/simple_chat)

SAP exposes a stateless HTTP endpoint POST /simple_chat that extensions can call for one-off AI tasks — translation, summarization, quick Q&A, code generation — without going through the WebSocket chat flow and without adding messages to the conversation history.

This is ideal when your extension needs a quick, single-turn AI call: translate text, summarize content, extract keywords, classify input, etc.

When to Use /simple_chat vs WebSocket

Feature	/simple_chat HTTP API	WebSocket (trigger_send_message)
Conversation history	❌ Stateless — no history	✅ Full chat history
Messages shown in UI	❌ Not added to chat	✅ Rendered in message list
Use case	One-off: translate, summarize, classify	Multi-turn chat, agent tasks
Response format	OpenAI-compatible JSON / NDJSON stream	messages_update / broadcast_messages events
Speed	Uses SAP's fast client config	Uses current active model provider

Endpoint

POST /simple_chat
Content-Type: application/json

The endpoint is on the same origin as the extension, so use a relative URL:

const res = await fetch('/simple_chat', { ... });

Request Format

{
  "messages": [
    { "role": "system", "content": "You are a professional translator." },
    { "role": "user", "content": "Translate 'Hello world' to Chinese." }
  ],
  "stream": false,
  "temperature": 0.7
}

Field	Type	Required	Description
messages	array	✅	Array of {role, content} objects (system/user/assistant)
stream	boolean	❌ (default false)	true for streaming, false for one-shot JSON response
temperature	number	❌ (default from settings)	0–2, lower = more deterministic

Non-Streaming Response (stream: false)

Returns a standard OpenAI-compatible ChatCompletion JSON object:

{
  "id": "chatcmpl-xxx",
  "object": "chat.completion",
  "created": 1234567890,
  "model": "gpt-4o",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "你好世界"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 20,
    "completion_tokens": 5,
    "total_tokens": 25
  }
}

Access the result: data.choices[0].message.content

Streaming Response (stream: true)

Returns NDJSON (one JSON object per line), matching OpenAI's streaming format. Each line contains a delta chunk:

{"id":"chatcmpl-xxx","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]}
{"id":"chatcmpl-xxx","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"你好"},"finish_reason":null}]}
{"id":"chatcmpl-xxx","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"世界"},"finish_reason":null}]}
{"id":"chatcmpl-xxx","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

Note: The stream does NOT send a [DONE] marker. Detect completion by checking choices[0].finish_reason.

JavaScript Usage Examples

Non-Streaming (Simple One-Shot Call)

/**
 * Call SAP's /simple_chat for a one-off AI task.
 * @param {Array} messages - [{role, content}, ...]
 * @param {number} [temperature=0.7]
 * @returns {Promise<object>} OpenAI-compatible ChatCompletion
 */
async function simpleChat(messages, temperature = 0.7) {
  const res = await fetch('/simple_chat', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ messages, stream: false, temperature })
  });
  if (!res.ok) {
    const err = await res.json();
    throw new Error(err.error?.message || `HTTP ${res.status}`);
  }
  return await res.json();
}

// ---------- Practical Examples ----------

// Translation
async function translate(text, targetLang = 'Chinese') {
  const res = await simpleChat([
    { role: 'system', content: `You are a translator. Translate to ${targetLang}. Reply ONLY with the translation, no explanations.` },
    { role: 'user', content: text }
  ]);
  return res.choices[0].message.content;
}

// Summarization
async function summarize(text, maxWords = 50) {
  const res = await simpleChat([
    { role: 'system', content: `Summarize in ≤${maxWords} words. Reply ONLY with the summary.` },
    { role: 'user', content: text }
  ]);
  return res.choices[0].message.content;
}

// Quick classification
async function classify(text, labels) {
  const res = await simpleChat([
    { role: 'system', content: `Classify into one of: ${labels.join(', ')}. Reply ONLY with the label.` },
    { role: 'user', content: text }
  ]);
  return res.choices[0].message.content.trim();
}

Streaming (Real-Time Display)

/**
 * Call /simple_chat with streaming. Yields delta content strings.
 * @param {Array} messages
 * @param {number} [temperature=0.7]
 * @returns {AsyncGenerator<string>} Yields delta content chunks
 */
async function* simpleChatStream(messages, temperature = 0.7) {
  const res = await fetch('/simple_chat', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ messages, stream: true, temperature })
  });
  if (!res.ok) {
    const err = await res.json();
    throw new Error(err.error?.message || `HTTP ${res.status}`);
  }

  const reader = res.body.getReader();
  const decoder = new TextDecoder();
  let buf = '';

  while (true) {
    const { done, value } = await reader.read();
    if (done) break;
    buf += decoder.decode(value, { stream: true });
    const lines = buf.split('\n');
    buf = lines.pop();  // keep incomplete line in buffer
    for (const line of lines) {
      if (!line.trim()) continue;
      try {
        const chunk = JSON.parse(line);
        const content = chunk.choices?.[0]?.delta?.content;
        if (content) yield content;
        if (chunk.choices?.[0]?.finish_reason === 'stop') return;
      } catch(e) { /* ignore parse errors for partial lines */ }
    }
  }
}

// Usage: render streaming response into an element
const el = document.getElementById('output');
el.textContent = '';
for await (const chunk of simpleChatStream([
  { role: 'user', content: 'Write a haiku about coding.' }
])) {
  el.textContent += chunk;
}

Error Handling

On error, the endpoint returns a JSON object with an error field:

{
  "error": {
    "message": "No model providers configured",
    "type": "server_error",
    "code": 500
  }
}

Always check res.ok and parse the error body.

Important Notes for /simple_chat

• Stateless: Each call is independent. No conversation context is preserved between calls.
• No UI impact: Results are NOT displayed in the main chat window. Your extension owns the rendering.
• Uses fast client: The endpoint uses SAP's "fast" model provider configuration. This may be a different model than the main chat.
• Same origin only: Extensions are served from the same origin, so no CORS issues. Use a relative URL (/simple_chat).
• Not a replacement for MCP tools: If you need the AI agent to call your extension, register MCP tools via WebSocket. /simple_chat is for your extension to call the AI, not the other way around.

---

Important Notes

• Extension ID format: {owner}_{repo} (e.g., heshengtao_sap-example)
• nodePort: 0 means auto-assign a free port (3100-13999 range)
• Always register beforeunload handler to send unregister_node_extension_mcp
• MCP works in both static and Node.js extensions — the register_node_extension_mcp message type name is historical; it works over WebSocket from any extension. Always follow the three-stage lifecycle: registerMcpTools() on WS open, handleMcpCall() on tool call, unregisterMcpTools() on beforeunload
• Font Awesome: Always use CDN (cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css). Relative paths like ../../fontawesome/ do NOT work for Node.js extensions (they're served from Express, not from SAP's static directory)
• Theme colors: Each extension defines its own identity via CSS variables on :root and body.dark. Do NOT force SAP's theme colors
• Always implement dark/light mode and Chinese/English i18n as basic functionality
• Transparent windows: Always implement compact mode. Without -webkit-app-region: drag, frameless windows cannot be moved. Without -webkit-app-region: no-drag on interactive elements, buttons become unclickable
• Close button: For transparent/frameless windows, the extension MUST provide its own close button since there's no native title bar

---

Reference Implementations

Study these real extensions for patterns:

• sap-lx-music — Static extension with MCP, transparent compact mode, dark/light theme, i18n, custom scheme invocation
• sap-example (heshengtao_sap-example) — Basic static chat UI extension
• sap-example-with-node (heshengtao_sap-example-with-node) — Node.js extension with Express backend

Resources

assets/

• assets/static-template/ — Complete starter template for static extensions
• assets/node-template/ — Complete starter template for Node.js extensions

references/

• references/package-json-spec.md — Complete package.json field reference
• references/node-entry-spec.md — Node.js entry point and lifecycle specification