Build a WhatsApp AI Agent with navin

This guide connects navin to WhatsApp through the whatsapp channel. The channel links as a WhatsApp device and uses the same navin agent runtime, tools, memory, and workspace as the CLI and WebUI.

What this guide builds

• WhatsApp optional dependencies installed
• a linked WhatsApp device session
• the whatsapp channel enabled in config.json
• one pairing-approved WhatsApp sender

Prerequisites

• A working local navin reply:

navin agent -m "Hello!"

• A WhatsApp account that can link a new device.
• A machine that can keep navin gateway running.

Install navin

python -m pip install navin-ai
navin webui   # configure provider & model in the platform (Settings → Providers)

Enable the WhatsApp channel

Install the optional channel dependency:

navin plugins enable whatsapp

Link WhatsApp as a device:

navin channels login whatsapp

Scan the QR code from WhatsApp -> Settings -> Linked Devices.

Merge this snippet into ~/.navin/config.json:

{
  "channels": {
    "whatsapp": {
      "enabled": true,
      "groupPolicy": "mention"
    }
  }
}

Omitting allowFrom enables pairing-only mode for private chats. groupPolicy defaults to "open" in the channel, but "mention" is safer for a first deployment.

Run navin gateway

navin channels status
navin gateway

Test a message

Send the bot a private WhatsApp message. It should return a pairing code. Approve it from a trusted local surface:

navin agent -m "/pairing approve ABCD-EFGH"

Send the message again after approval. The reply should use the same model and workspace as your local CLI check.

Security notes

• Treat the WhatsApp session database as account access.
• Prefer pairing-only mode for first setup. Add allowFrom only when you want a static allowlist.
• Keep groupPolicy as "mention" before adding the bot to groups.
• Avoid allowFrom: ["*"] unless the bot is intentionally public or isolated.

Troubleshooting

• If QR linking fails, rerun navin channels login whatsapp.
• If you are migrating from the old bridge, remove bridgeUrl and bridgeToken, then re-login.
• If a sender appears as a LID instead of a phone number, let navin learn the mapping at runtime or use lidMappings in the full reference.
• If a first private message returns a pairing code, approve it before testing normal replies.

Next: memory, automations, MCP tools

• Secure local AI agent