What if the browser task is on a protected site?

The OpenClaw browser hits the same anti-bot walls as any other stock Chromium when targeting Cloudflare-, DataDome-, or PerimeterX-protected sites. The fix is to set the active browser profile to remote and point it at a stealth backend like Human Browser, which runs a stealth-patched Chromium on residential IPs. Your chat experience and the skill code don't change; the browser underneath does.

Can OpenClaw handle group chats on WhatsApp / Telegram?

Yes, with configuration. In groups the agent only responds when explicitly mentioned (@bot or @openclaw) by default, to avoid taking over every conversation. You can adjust the trigger pattern per channel in the Gateway config — for example, always respond in a private group, never respond in a public group, respond only to messages from specific user IDs.

June 15, 2026 · 9 min read · AI Agents · OpenClaw

OpenClaw on WhatsApp & Telegram (2026): Browser Tasks From Any Chat

Q: Can OpenClaw use WhatsApp?

Yes. WhatsApp is one of 20+ built-in channels OpenClaw supports. After install, openclaw channel add whatsapp opens a QR-code pairing flow — scan it on your phone (WhatsApp → Linked Devices → Link a Device) and any message you send to that paired number routes to the OpenClaw agent. The pairing stays linked until you remove the device on your phone, the same way WhatsApp Web works.

Q: Can OpenClaw use Telegram?

Yes — both bot accounts and personal accounts. The standard setup is via a Telegram bot: create one with @BotFather to get a bot token, then run openclaw channel add telegram and paste the token. The agent answers any direct message to that bot. Personal Telegram via the Zalo / WeChat / Telegram Personal class of channels is also supported for the case where you want the agent to use your own account instead of a separate bot.

Q: Is it safe to give the agent access to my chat apps?

OpenClaw runs locally — the Gateway process is on your hardware, the chat-app sessions are paired to that local process, and no message data is sent to OpenClaw's servers. The model provider (Anthropic, OpenAI, etc.) sees the prompts because it has to generate responses; that's an honest trade-off with any LLM. You can mitigate by routing sensitive tasks to a local model via llama.cpp.

Q: Can the agent run browser tasks from a chat message?

Yes — that's a primary use case. Message 'open my Linear board and tell me what's overdue' from WhatsApp; the OpenClaw agent routes the task to its browser tool (or the Linear MCP server) and responds in the same chat. The bundled browser-automation skill handles snapshot, stable-tab, stale-ref recovery so multi-step flows don't fall over on the third re-render.

A chat bubble on a phone routing to a central automation hub that controls floating browser windows running tasks — OpenClaw on WhatsApp and Telegram illustration

If you searched openclaw whatsapp or openclaw telegram, you're probably trying to answer two questions: "how do I pair this?" and "is it actually useful in practice?" This post covers both. We'll wire both channels end to end, walk through the kinds of tasks that work cleanly from chat, and call out where the local OpenClaw browser hits a wall and what to do about it.

Disclosure: we make Human Browser, which is the stealth backend we'll wire in for protected-site browser work. The non-protected examples don't need it.

Prerequisites
Pairing WhatsApp
Pairing Telegram
First test message
Browser tasks that work cleanly from chat
When the target is protected
Group chats and triggers
FAQ

Prerequisites

You need an OpenClaw install with the browser-automation skill enabled. If you don't have one yet, the install path is:

$ curl -fsSL https://openclaw.ai/install.sh | bash
$ openclaw onboard      # walks through provider + model + workdir
$ openclaw skill add browser-automation
$ openclaw doctor       # should be green

The full step-by-step is in our companion install guide. From here we assume openclaw --version works.

Pairing WhatsApp

One command, two minutes:

$ openclaw channel add whatsapp

  Pairing WhatsApp...
  ┌──────────────────────────┐
  │     [QR code displays]    │
  └──────────────────────────┘
  ▷ On your phone: WhatsApp → Settings → Linked Devices →
    Link a Device → scan this QR.

  ✓ paired as "OpenClaw on iPhone — Ilya"
  ✓ welcome message sent to your phone

From your phone's WhatsApp, you should see a new linked-devices entry called something like "OpenClaw — Ilya". Messages to your own number now route to the OpenClaw agent the same way WhatsApp Web works.

Pairing Telegram

Two minutes via a bot token:

Open Telegram, message @BotFather, run /newbot, follow the prompts (name + username). You'll get a bot token that looks like 12345:ABC-DEF....
Paste it into OpenClaw:

$ openclaw channel add telegram

  Bot token (paste from @BotFather): 12345:ABC-DEF...

  ✓ Telegram bot @YourOpenClaw verified
  ✓ updates webhook registered
  ✓ welcome message sent

From any Telegram client, message your new bot — anything works as a first test. The agent responds in the same chat. If you prefer your personal Telegram account over a bot, the telegram-personal channel plugin is the alternative; it uses MTProto auth (phone code) instead of a bot token.

First test message

Send "What can you do?" from either channel. The agent should reply with a one-paragraph summary of the model it's using, the channels it can hear on, and the skills installed in the active workdir. If that round-trip works, the chain is healthy and you're ready for real tasks.

If the agent doesn't respond, run openclaw doctor --deep from the Gateway machine. It will pinpoint which link broke — model provider unreachable, channel auth expired, skill load error, etc.

Browser tasks that work cleanly from chat

The interesting part of WhatsApp / Telegram + OpenClaw isn't the chat — it's that any message can trigger a real browser action. These are the patterns we see work well in the wild:

1. "Check this dashboard and tell me what changed"

The agent opens the URL via the bundled browser tool, takes a snapshot, compares it to the previous snapshot (held in memory across sessions), and reports deltas. Works for internal dashboards, status pages, and any HTML page you'd glance at on the way home.

2. "Summarise the top 10 entries on hacker news right now"

Browser tool fetches the page, the agent extracts and ranks. Three-paragraph reply in the chat. The browser-automation skill handles the "page didn't finish loading" case so the summary isn't partial.

3. "File a ticket in our help-desk: subject X, body Y, assignee Z"

The agent opens the help-desk UI, fills the new-ticket form, submits, and replies with the ticket URL. The first time, you teach it the URL and any UI quirks via a quick skill; from then on it's reliable.

4. "Reschedule the 3pm meeting to tomorrow at 10am"

For Google / Outlook Calendar, the cleanest path is an MCP server (calendar APIs are far more stable than UIs). But the browser-tool path also works if you'd rather not introduce API keys.

A phone displaying a WhatsApp chat bubble on one side and a browser window showing a flight-booking form on the other side, joined by an arrow

When the target is protected

The local OpenClaw browser hits the same walls every stock Chromium does. The first time you ask "open this Cloudflare-protected site from my phone," you'll see the agent come back with "blocked by an interstitial." The fix isn't in the skill or the chat plumbing — it's in the browser layer.

OpenClaw's browser tool supports a remote profile that connects to an external Chrome over CDP. Wire it to Human Browser — a stealth-patched Chromium on a residential IP, with a CAPTCHA solver — and the same chat-triggered tasks reach the protected target without changing anything else:

# one-time, on the Gateway host:
$ export HB_CDP_ENDPOINT="wss://agent.humanbrowser.cloud/cdp?token=hb_live_..."
$ export OC_BROWSER_REMOTE_CDP="$HB_CDP_ENDPOINT"
$ openclaw browser --browser-profile remote start

# from WhatsApp / Telegram:
#   "go to <protected target> and tell me the headline news"
# the chat-side experience is unchanged; the browser underneath now
# runs on a residential IP with patched CDP fingerprints.

The full integration recipe is in our remote-CDP companion post. For chat-triggered scraping where you don't want to micro-manage the browser layer, the MCP-based wiring is equivalent and slightly cleaner — covered in our OpenClaw + MCP piece.

Want a stealth backend for chat-triggered browser tasks?

Human Browser exposes a remote CDP endpoint (and an MCP server) that drops into OpenClaw and routes through residential IPs with stealth Chromium and CAPTCHA solving. Pay-as-you-go: $0.10/hr browser time, $4/GB residential proxy, $0.005/captcha, $0.02 per agent step with the AI included. No subscription, unused balance never expires. Try it for $1 — humanbrowser.cloud.

Group chats and triggers

In a WhatsApp or Telegram group, the agent won't respond to every message by default — it'd be unbearable. The trigger patterns you can configure in the Gateway:

Mention — the agent responds when someone tags it (@openclaw or the bot username on Telegram). This is the default.
Prefix — only messages starting with a configured prefix (e.g. !ai) trigger a response.
Always — every message in the group goes to the agent. Useful for a 1:1 chat that happens to be a group; almost always wrong for real groups.
Allowlist — only messages from specific user IDs trigger. Useful for "the agent only listens to me in this shared group."

Configure per-channel via openclaw config edit under channels.<name>.trigger. Restart the Gateway (openclaw restart) to pick up the change.

FAQ

Can OpenClaw use WhatsApp?

Yes — built-in channel. openclaw channel add whatsapp opens a QR pairing flow same as WhatsApp Web.

Can OpenClaw use Telegram?

Yes — both bot accounts (via @BotFather token) and personal accounts (via the telegram-personal channel plugin and MTProto auth).

Is it safe to give the agent access to my chat apps?

The Gateway runs locally; chat sessions pair to your machine. The model provider sees prompts (necessary for response generation). Sensitive tasks can be routed to a local llama.cpp model to keep prompts on-device.

Can the agent run browser tasks from a chat message?

Yes — main use case. Dashboards, summarisation, ticketing, scraping, posting. The bundled browser-automation skill makes multi-step flows survive DOM drift.

What about protected sites?

The local OpenClaw browser hits the same walls as any stock Chromium. Wire the remote browser profile to a stealth backend like Human Browser and the chat-triggered tasks reach Cloudflare- and DataDome-protected targets unchanged.

Can OpenClaw handle group chats?

Yes — configurable triggers (mention, prefix, always, allowlist). Mention-only is the safe default.

OpenClaw Install Guide (2026) — Tutorial
OpenClaw Browser: The Tool, the Remote-CDP Mode, and When to Pair It With Human Browser (2026) — AI Agents
OpenClaw and MCP: Connecting Your Personal Assistant to External Tools (2026) — AI Agents
OpenClaw Skills & ClawHub: The Plugin Marketplace for Agent Capabilities (2026) — AI Agents