WhatsApp MCP servers now range from weekend scripts to hosted products. They differ on five axes that matter, and none of those axes is the number of tools.
The five axes
- Session durability — does it survive a dropped link without you re-scanning a QR code?
- Auth model — per-client revocable tokens, or one shared secret in a config file?
- Send guardrails — are bulk limits enforced in the server, or suggested in a prompt?
- Context quality — does search return the surrounding conversation, or isolated matching lines?
- Isolation — is your data in its own environment, or a shared database with a tenant column?
What the self-hosted scripts get right
Cost and control. If you are comfortable operating it, a local server keeps everything on hardware you own, and the tool surface is exactly what you wrote.
What they usually get wrong
- A single long-lived token pasted into several clients, revocable only by rotating it everywhere.
- No enforcement of per-chat permissions, because there is no permission model.
- Sends with no cap, which is how a well-meaning agent gets your number flagged.
- Silent session loss.
Before you connect any of them
Run the adversarial test: ask for a forty-person broadcast, a read of a blocked chat, and a bulk delete. A server worth trusting refuses all three, tells you why, and shows you the refusal in an audit trail you can read.
KKHS Assist exposes its MCP endpoint from the dashboard with two scopes, per-client tokens, and every send subject to the same per-chat rules as the assistant itself.