whatsapp

Files (7)

• SKILL.md
• SKILL.md · persona/
• send.py · scripts/
• package-lock.json · server/
• package.json · server/
• server.js · server/
• start.sh · server/

SKILL.md (first lines)

# WhatsApp

Personal WhatsApp assistant that connects to your own number via Baileys (multi-device). Responds to messages using the full Claude CLI — no separate API key needed. Protects your number from strangers, manages groups via commands, and remembers context across restarts via Engram.

---

## Setup

### 1. Install dependencies

```bash
cd skills/communications/whatsapp/server
npm install
```

### 2. Configure persona

Create `~/.dexter/whatsapp-persona.json`:

```json
{
  "name": "Your Name",
  "about": "Freelance developer, React and Node.js specialist.",
  "tone": "casual and friendly",
  "language": "es",
  "allowFrom": ["+521XXXXXXXXXX"],
  "stranger_reply": "Hey, I can't chat right now. I'll get back to you soon 👋",
  "llm": { "provider": "anthropic", "model": "claude-haiku-4-5-20251001" }
}
```

| Field | Description |
|-------|-------------|
| `allowFrom` | Your phone number(s) — get full AI responses. International format with `+`. |
| `stranger_reply` | Fixed reply for unknown numbers. If omitted, strangers get an LLM response via `callLLM`. |
| `wake_word` | Word required to trigger Dexter (default: `"dexter"`). Set to `null` to disable. |
| `allowedGroups` | Group JIDs where Dexter responds. Managed via `dexter join` / `dexter leave` commands. |
| `groups` | Per-group config: personality, standing instructions, allowed members. Managed via commands. |
| `llm.provider` | LLM for stranger responses: `"anthropic"`, `"openai"`, or `"ollama"`. |

### 3. Start the server

#### QR pairing (simplest)

```bash
node skills/communications/whatsapp/server/server.js --qr
```

A QR code is printed in the terminal. Scan with WhatsApp → ⋮ → Linked devices → Link a device.

#### Phone number pairing (headless)

```bash
WA_PHONE=+521XXXXXXXXXX node skills/communications/whatsapp/server/server.js
```

A pairing code is shown. In WhatsApp → ⋮ → Linked devices → Link with phone number → enter the code.

Credentials are saved in `~/.dexter/whatsapp/` — subsequent starts reconnect automatically without scanning.

#### Keep it alive with pm2 (recommended)

```bash
npm install -g pm2
WA_PHONE=+521XXXXXXXXXX pm2 start skills/communications/whatsapp/server/server.js --name dexter-wa
pm2 save       # persist across reboots
pm2 startup    # auto-start on system boot
```

```bash
pm2 status            # check if running
pm2 logs dexter-wa    # tail logs
pm2 restart dexter-wa
pm2 stop dexter-wa
```

---

## How it works

### Access tiers

| Sender | Behavior |
|--------|----------|
| Numbers in `allowFrom` | Full Dexter AI response — spawns `claude -p` with complete tool access |
| Self-chat (own number) | Same as above — write to yourself, Dexter responds |
| Unknown numbers | Ignored unless they say the wake word — then `stranger_reply` or LLM persona |
| Groups (not enabled) | Fully ignored |
| Groups (enabled) — owner | Full unrestricted Dexter response, complete tool access, owner instructions applied |
| Groups (enabled) — allowed member | Same as owner in that group (granted via `dexter allow`) |
| Groups (enabled) — non-owner | Restricted: general knowledge only, no machine/filesystem/code access |

### Phone number matching

Numbers are matched by **last 10 digits** — country code variants are handled automatically. Mexico's `+5283...` and WhatsApp's internal `+52183...` both match the same entry. Use whatever format is printed on your SIM.

### Self-chat

Writing to your own number is the primary interaction mode. Messages sent to yourself (`fromMe: true`) are processed normally — Dexter responds in that same conversation.

### AI responses (owner)

When the owner messages, Dexter spawns `claude -p "message"` from the Dexter project root. Claude has full access to `DEXTER.md`, `CLAUDE.md`, Engram memory, and all skills.

To use a different CLI:
```bash
DEXTER_AGENT=opencode node server.js
```

---

## Memory — Engram integration

If `engram` is installed and available in PATH, Dexter uses it for persistent cross-session memory. If not, it falls back to in-process RAM history (lost on restart).

### How it works

Instead of manually tracking conversation history, Claude receives a system prompt that instructs it to use Engram tools directly:

From EmilioUgarte65/Dexter, Apache-2.0.

More from EmilioUgarte65/Dexter