Intercom Support Channel

This skill sets up @othreecodes/openclaw-intercom — an OpenClaw channel plugin where customers message any surface connected to Intercom (WhatsApp, Instagram, Facebook, in-app Messenger, SMS, email) and the agent answers them directly: greets, resolves, tags, annotates, escalates to a human when needed, then stays out of the human's way.

This is different from Intercom API skills, which help an operator list conversations or draft replies on request. This makes the agent the support teammate customers talk to.

What the plugin handles for you

• Hybrid inbound (polling + webhooks) with crash-safe dedupe — no double replies
• One agent session per conversation; message bursts coalesce into one coherent reply
• Reads customer screenshots (uploads and Instagram inline images) via image understanding
• Transcribes voice notes — WhatsApp and Instagram audio reaches the agent as text instead of "the customer attached a file"
• Replies render as real HTML: numbered steps, bullets, bold
• Inline actions parsed from the agent's reply: [[close]], [[escalate: reason]], [[note: text]], [[tag: label]]
• Escalation hands the conversation back to the inbox it came from and permanently mutes the agent on it
• Human ownership is absolute: once any teammate replies in a conversation, the agent takes no further turn in it, even on new customer messages. On a shared team inbox a reopened conversation drops back to the queue, and without this the agent walks into threads a human is actively working
• Session labels name the channel a conversation came in on (Ada (WhatsApp)), so a multi-channel inbox stays readable
• First run absorbs the existing backlog instead of answering the whole inbox
• Channel scoping: pilot on one surface (allowedChannels: ["instagram"]), widen later

Setup

1. Install the plugin — pin the version so you get exactly the release you reviewed (this skill and the plugin share the same publisher, and the package is source-linked on ClawHub, so every version maps to a verifiable GitHub commit):

   openclaw plugins install clawhub:@othreecodes/openclaw-intercom@1.0.12
   

   Check ClawHub for the latest version and its changelog before upgrading; treat plugin upgrades like any other production dependency change.

2. Get authorization first. This agent will reply to real customers under a real teammate identity — confirm with whoever owns the support workspace before enabling it.

3. In the Intercom Developer Hub, create a dedicated app for the bot (don't reuse a broader app's token), copy its Access token, and grant only what you'll use: read/write conversations, plus read contacts and read/write tags if you want those features. Keep the token rotatable — you'll want revocation to be one click if anything looks wrong.

4. Configure the channel (~/.openclaw/openclaw.json, or the OpenClaw dashboard's channel settings, where every option renders as a form):

   {
     "channels": {
       "intercom": {
         "enabled": true,
         "token": "YOUR_ACCESS_TOKEN",
         "inbound": "both",
         "webhookSecret": "YOUR_APP_CLIENT_SECRET",
         "allowedChannels": ["instagram"]
       }
     }
   }
   

   token is the only required field. For webhooks, subscribe the app to conversation.user.created and conversation.user.replied, pointed at https://YOUR_GATEWAY/intercom/webhook — the signing secret is the app's client secret.

5. Restart the gateway. On first run the plugin logs how many existing conversations it absorbed without replying; from the next customer message onward the agent answers.

Roll out in stages. Start with allowedChannels scoped to one low-stakes pilot channel and escalationAssigneeId pointed at a real human team before the first customer message — an autonomous agent without a working hand-off path is not ready for production. Watch its first days of conversations, then widen the channel list as confidence grows.

Give the agent a support persona and grounding rules in its workspace AGENTS.md — the plugin carries the messages; the agent's quality comes from its instructions.

Two things worth writing into those instructions before you go live, both learned the hard way on a production inbox:

• Never let the agent identify which app a screenshot belongs to. It will get it wrong in both directions — naming a competitor's app, or telling a customer their own genuine screen is fake. Have it describe what it sees, believe the customer about their own screen, and ask rather than assert.
• Scope it explicitly to your product. Without a boundary it will cheerfully answer unrelated questions, which is both off-brand and the same weakness an injected instruction exploits.

Full documentation

• Package: clawhub:@othreecodes/openclaw-intercom (ClawHub → Packages)
• Source, README and config reference: https://github.com/othreecodes/openclaw-intercom