Muhammad Kashif

Engineering notes / Muhammad Kashif

OpenClaw Over Telegram and WhatsApp: Agents That Message Back

OpenClaw over Telegram and WhatsApp turned the agent from something I babysit in a desktop window into something that messages back while I am away from the keyboard. Portfolio sites, narrative games, and match-3 prototy…

OpenClaw Over Telegram and WhatsApp: Agents That Message Back

OpenClaw Over Telegram and WhatsApp: Agents That Message Back

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

OpenClaw over Telegram and WhatsApp turned the agent from something I babysit in a desktop window into something that messages back while I am away from the keyboard. Portfolio sites, narrative games, and match-3 prototypes from prompts proved the hands still work when a tool is missing and the agent switches to Python. Messaging channels made the threat model louder: a chat bubble is an easy place to accidentally grant production powers.

Chat as a remote control, not a root shell

Telegram and WhatsApp are brilliant for enqueueing tasks and receiving summaries. They are terrible as the only authentication factor for destructive actions. I bind the bot to a allowlisted phone number, require a second factor for mutates, and keep a desktop confirmation for anything that spends money or sends email as me.

Remote control means I can steer. It does not mean the agent can drive off without me when a spoofed message arrives at 1 a.m.

Adaptation when tools go missing

The notes showed an agent switching to Python when a tool was missing, still delivering a portfolio site, a narrative game, and a match-3 game from prompts. Adaptation is valuable. Silent adaptation that installs packages or opens network paths I did not approve is not.

I require the agent to announce the fallback before executing it. If it cannot announce on the same channel, it waits. Cleverness without notice is how supply-chain surprises land in my home directory.

What belongs on WhatsApp

Status pings, queue lengths, links to draft previews, reminders. Not API keys. Not customer dumps. Not "here is the .env." I have seen people paste secrets into chat because it felt faster than a vault. Faster until the backup of the chat becomes the breach narrative.

OpenClaw get summaries. Secrets stay in the vault. Previews stay behind expiring links.

Telegram bots with human cadence

I rate-limit outbound messages so the agent cannot spam me into approving without reading. Burst alerts go to a quiet channel; approvals live in a loud one I only open when working. Fatigue is an attacker. Design for fatigue.

Commands are explicit verbs: research, draft, status, stop. Free-form natural language that might mean delete is disallowed for high-risk tools.

Games and portfolios as sandbox proofs

Prompt-built games and portfolio sites are how I test channel latency and tool fallbacks without client data. When the sandbox looks good, I still do not promote the same bot token to production. New environment, new token, new allowlist.

Promotion theatre — reusing demo tokens — is a classic own-goal. I write the new token step into the runbook so tired-me cannot skip it.

Cross-channel identity

The same agent on Telegram and WhatsApp must not mean the same privileges. I treat channels as separate personas with separate scopes, echoing the Agents of Chaos lesson about spoofed identities. A compromised WhatsApp session should not inherit Telegram's send rights to customers.

Revocation drills monthly: kill a channel token and confirm the other channel still respects its own narrower world.

What I sell

Messaging remote-controls for research and drafting agents, with approval UX that survives a busy commute. I do not sell fully autonomous customer messaging on day one. Day one is inbound commands from me. Outbound to customers comes after logs are boring.

Rule on the lock screen

Chat can wake me. Chat cannot spend. Chat cannot send as me to the outside world without a typed approve. OpenClaw may message back. It may not become me.

Channel revoke drill

Monthly I revoke the Telegram token and confirm WhatsApp still respects its narrower scopes — then reverse. OpenClaw remote control that cannot survive a revoke drill is a single point of failure with chat bubbles.

I also rotate the allowlisted phone numbers after contractor changes. Stale allowlists are how ex-teammates keep waking agents.

Outbound customer mute

Default: messaging agents cannot outbound to customers. Enabling outbound requires a written exception with sample scripts and a human on-call. Telegram convenience is not a CRM strategy.

Command verb list

research, draft, status, stop — and nothing that sounds like delete or pay. Free-form natural language for high-risk tools is disabled on WhatsApp and Telegram alike. OpenClaw adapts tools; it does not invent verbs.