How Do You Build an AI Agent for Intercom in 2026? API vs. Connector

You can build an AI agent for Intercom two ways: write it yourself against the Intercom REST API (current version 2.16, released 15 July 2026), which takes weeks once webhooks, retries and evaluation are included, or connect a platform that already has an Intercom connector and test an agent on real conversations the same day.

What does an Intercom AI agent actually do?

The loop is the same as any support agent, wired to Intercom: a new conversation arrives, the agent reads it and the contact's history, it calls tools (look up the order, search your help center, check a policy), and it posts a reply on the conversation or hands it off to a human. The hard part is doing that reliably against real conversations. (New to agents generally? Start with AI agents for customer service.)

How do you build one against the Intercom API?

Intercom has a full REST API, so you can build the agent yourself. The pieces:

• Read conversations: the Conversations API pulls a conversation, its parts and the contact.
• Reply or act: post an admin reply to the conversation, or snooze, close or assign it through the API.
• Trigger on new conversations: an Intercom webhook (topics such as conversation.user.created and conversation.user.replied) calls your endpoint when a conversation is created or a customer replies.
• Pin a version: every request sends an Intercom-Version header. The API changelog lists 2.16 as the newest numbered version.

A reply tool against the real Intercom API looks like this:

import requests
BASE = "https://api.intercom.io"
HEADERS = {
    "Authorization": f"Bearer {ACCESS_TOKEN}",
    "Intercom-Version": "2.16",
    "Content-Type": "application/json",
}

def reply_to_conversation(conversation_id: str, admin_id: str, body: str):
    r = requests.post(f"{BASE}/conversations/{conversation_id}/reply",
        json={"message_type": "comment", "type": "admin",
              "admin_id": admin_id, "body": body},
        headers=HEADERS, timeout=10)
    r.raise_for_status()          # handle 429 rate limits, token refresh
    return {"status": "sent"}

Wrap that (and a get_contact and a search_help_center) as tools in an agent loop on the model of your choice. Our from-scratch walkthrough is in how to build an AI agent: from scratch vs. platform.

What does the production part involve?

The API calls are the easy bit. An Intercom agent you would trust also needs:

• A webhook endpoint that answers fast. Intercom gives you 5 seconds to respond. Miss it and the notification is treated as failed and retried once after 1 minute, so the endpoint should acknowledge first and run the agent in a background job.
• Signature checks and dedupe. Intercom signs each notification with an X-Hub-Signature header: sha1= plus an HMAC-SHA1 of the body keyed with your app's client secret. Store event IDs so a retried event never produces a second reply.
• Rate-limit handling. Intercom's default is 10,000 API calls per minute per app and 25,000 per minute per workspace, shared across every app installed there.
• Grounding in your Help Center (embed the articles and keep them fresh), guardrails (escalate refunds, never leak PII), 24/7 hosting, observability, and an eval harness over real historical conversations.

That is weeks of work plus upkeep, none of it specific to your business.

How does connecting Intercom through Macha work?

If the goal is a working agent on your Intercom rather than a project to maintain, a platform with a native connector is much faster. Macha layers on top of Intercom. It is not an Intercom replacement or a Messenger app from the App Store; it connects through OAuth and works autonomously:

• Connect once: authorize Intercom and the agent can get and search conversations, look up contacts, reply publicly or add an internal note, assign, tag and close.
• Trigger on the right events: built-in triggers fire on the first customer message, on every customer message, on assignment and on close, so you do not host a webhook endpoint.
• Ground on your Help Center: the Search Help Center Articles tool lets replies quote your real Intercom articles.
• Any other system as a tool: Custom Tools turn your order or billing APIs into agent capabilities by describing them.
• Runs and grades: Agent Analytics show every run, and simulations replay past conversations against the agent before it goes live.

Custom Tools are how you reach the systems the help desk does not cover: anything with a REST endpoint becomes an agent capability.

You don't build the agent yourself: the Macha team analyzes your past conversations, builds the knowledge base and the agent, and runs it in safe mode, with drafts as internal notes in Intercom, until the drafts are right. Pricing is one plan set by ticket volume, from $299 a month for up to 750 tickets (about $0.40 a ticket), billed per ticket rather than per message, and it includes setup and a dedicated success manager who handles your changes. The same agents run on Zendesk, Freshdesk, Gorgias, Front, HubSpot or Intercom.

So which should you build?

Build on the API if you need a fully custom runtime, have strict data-residency needs, or the agent is your product. For most Intercom teams who want conversations resolved, a native connector gets a measurable agent onto your real conversations much faster. You can start free on Macha with $50 of usage and no credit card, connect Intercom, and the Macha team builds the first agent and tests it on your past conversations before it replies to anyone.

FAQ

Is an Intercom AI agent the same as Intercom's own Fin? Not necessarily. Fin is Intercom's built-in agent at $0.99 per outcome, and Salesforce completed its acquisition of Fin on 10 September 2026. You can also run a model-agnostic agent on top of Intercom (through its API or a platform like Macha), which helps if you want a specific model, custom tools, or to grade the agent your own way.

Do I need to install an Intercom Messenger app? No. Macha connects to Intercom through OAuth and works autonomously; it is an AI layer on top of Intercom, not an App Store app you install into the Messenger.

Can the agent take actions, not just answer? Yes. Through the Intercom API it can reply, snooze, assign and tag, and your own systems' APIs become tools for order lookups or refunds, with guardrails on what it can do unattended.

How do I make sure it's accurate before it replies to customers? Ground it in your Help Center and replay it against real historical conversations first, start in a draft-and-approve mode, and widen autonomy per conversation type as it earns trust.