What we do
Tools
Industries
Department
How we workLet's talk →
Home/Intercom/Intercom MCP
Intercom MCP server

Intercom MCP: Point Your AI Client at Your Workspace

Intercom runs a hosted MCP server, so Claude, Cursor, VS Code or an agent you wrote can search conversations, look up contacts and draft Help Center articles without a custom integration. Below: the URL for your region, the auth choice, what each tool does, and how this differs from giving Fin its own MCP connectors.

Endpoint, auth, region

US workspace: https://mcp.intercom.com/mcp. EU workspace: https://mcp.eu.intercom.com/mcp. Both use Streamable HTTP; the older /sse endpoints are deprecated.

Auth: OAuth in the browser is the recommended route. A bearer API token works too, but skips the consent step and doesn't get the Fin Operator tools.

Australia-hosted workspaces aren't supported yet.

From a blank client config to a first search

  1. 1

    Add the server to your client

    In Claude, Claude Code, Cursor, Windsurf or VS Code, add a remote MCP server with your region's URL. Intercom supports Dynamic Client Registration, so most clients register themselves when you connect.

    Watch for: OAuth callbacks on localhost or 127.0.0.1 work on any port. Any other redirect URI has to be allowlisted first.

  2. 2

    Sign in and approve the permissions

    You'll see a consent page, then Intercom's own authorization screen. The connection asks to read users and companies, read and write conversations, read and write articles, and use Fin Operator.

  3. 3

    Run a read-only call first

    search_conversations and get_conversation pull threads; search_contacts, get_contact, list_companies and get_company cover people and accounts. The universal search tool takes a query that must name an object_type, and fetch returns the full record by ID.

  4. 4

    Decide which write tools anyone should use

    Writes are limited: add_internal_note on a conversation, and create_article or update_article in the Help Center. With Fin Operator and OAuth, five more tools appear, including apply_proposals, which changes your live Fin Operator configuration.

    Watch for: an internal note on a snoozed conversation assigned to someone else reopens it. Notes never reach the customer, but they do land in a teammate's queue.

Two different things both get called "Intercom MCP"

Intercom as the server

An outside AI reads Intercom

  • Claude, Cursor or your own agent connects to mcp.intercom.com.
  • It reads conversations, contacts, companies and articles.
  • It writes internal notes and articles, nothing a customer sees.
  • Intercom API rate limits apply to every call.
Fin as the client

Fin reads your other tools

  • Added in Fin as MCP connectors, from templates (Stripe, Shopify, Linear, Zapier) or a custom server.
  • Each tool from the server becomes a Data connector you configure before Fin uses it.
  • Connects with OAuth or an access token.
  • Custom MCP connectors are available to US-hosted workspaces only.

If your question is "how does my agent see Intercom", use the hosted server. If it's "how does Fin see my order system", you need an MCP server in front of that system, and for most in-house databases nobody has written one yet.

Settle these before an agent touches the production workspace

  • Whose token is it?

    A bearer token acts with whatever access that token has, for every person using the client. OAuth connects as the person who signed in and approved it.

  • Should this agent write at all?

    A research or reporting agent only needs the search and get tools. Strip the article and note tools from its instructions, or give it a client that only exposes the reads.

  • How many calls will one question cost?

    An agent summarizing a month of conversations can make hundreds of get_conversation calls. They count against the same rate limits as your other Intercom integrations.

  • Where does the answer go?

    Conversation text can hold personal data. Know which model provider and which logs see it before you point a client at real customer threads, and check that EU data stays on the EU endpoint.

MCP and Intercom, the questions that come up first

What is MCP?

The Model Context Protocol, an open standard for connecting AI assistants to tools and data. A system publishes an MCP server once, and any MCP-compatible client can call it, so you don't build a new integration for every AI tool and every system it needs to reach.

Does Intercom support MCP?

Yes, in two directions. Intercom hosts an MCP server that outside AI clients connect to, and Fin can act as an MCP client, calling other systems through MCP connectors during a conversation.

Is MCP the same as a Custom Action in Intercom?

No. A Custom Action is a single API call Fin makes for one defined purpose, which you configure by hand. MCP is a general protocol: one server can offer many tools, and in Fin each of those tools becomes a Data connector. MCP isn't limited to Intercom's own framework, which is why the same server can also serve Claude or Cursor.

Do I need to understand MCP to use Fin effectively?

No. Plenty of teams run Fin well on its content, its built-in connectors and a few Custom Actions. MCP matters once you connect a separate agent to Intercom data, or want Fin to call a system that already has an MCP server.

Can the Intercom MCP server reply to customers?

No. It can add internal notes, which only teammates see, and create or update Help Center articles. It has no tool for sending a customer-facing reply.

Which AI clients work with it?

Intercom lists Claude Desktop, Claude.ai, Claude Code, OpenAI, Cursor, Windsurf and VS Code, and any client you build yourself. MCP Inspector is handy for testing a connection before you wire it into anything.

Let's talk

Tell us the one thing your team does manually that eats up time. We read every message and reply within a day.