Intercom MCP Server

The Intercom MCP Server implements the Model Context Protocol and opens up Intercom customer data to external AI systems, letting teams build their own AI assistants, internal tools, or automations without wiring up the raw Intercom REST API from scratch. Per the official announcement from 2025-05-01, the server runs on Cloudflare's infrastructure and is meant to remove the technical complexity of a direct integration, so even non-technical teams can set up AI integrations without writing code. The technical documentation lives at https://developers.intercom.com/docs/guides/mcp.

Distinguishing the MCP server from Fin as an MCP client

Intercom uses the Model Context Protocol in two directions that are easy to confuse. The Intercom MCP Server described here exposes Intercom data to any external AI client, for example a custom-built support bot or an internal analytics tool. Separately, Fin, Intercom's own AI support agent, can itself act as an MCP client and connect to third-party MCP servers such as Stripe, Shopify, or Linear to pull data from those systems into its own answers. This catalog entry covers only the first direction: the MCP server through which external tools access an Intercom account.

Tool catalog

According to official documentation, the server exposes 14 tools across four groups. Two universal tools (search, fetch) provide cross-cutting search over conversations and contacts plus detail retrieval for a found object. Six direct API tools cover conversations and company data: search_conversations, get_conversation, search_contacts, get_contact, list_companies, get_company. Five tools revolve around the help center: list_articles, search_articles, get_article, and the write tools create_article and update_article. One further tool, add_internal_note, adds an internal note visible only to the team to a conversation.

Setup and authentication

The server is available exclusively as a hosted remote service; according to official documentation, there is no local installation path or open-source repository. Depending on region, a dedicated endpoint is used: https://mcp.intercom.com/mcp for US workspaces and https://mcp.eu.intercom.com/mcp for EU workspaces. Two authentication paths are available: OAuth with a browser-based consent screen (recommended by Intercom) or a direct bearer token. In MCP clients without native remote support, the connection can be established via the mcp-remote helper tool.

Regional availability

Per the provider, the server is available for US- and EU-hosted Intercom workspaces; AU-hosted workspaces are not currently supported. Requests to the EU endpoint are processed within the EU according to the documentation, in line with Intercom's regional data hosting commitments. Before setup, it's worth checking which region your own Intercom workspace is hosted in to use the correct endpoint.

Security and write access

Alongside pure read tools, the server also offers write access: help center articles can be created and updated, and internal team notes can be added to ongoing conversations. The permissions actually usable depend on the scopes granted during OAuth consent or the permission scope of the bearer token used — documented scopes include "Read conversations," "Write conversations," "Read and list users and companies," and "Read and write articles." Because customer conversations and article content feed into a connected AI model's responses, access rights should be limited to what's actually needed, and bearer tokens should be stored securely like any other API key. The documentation also notes that regular Intercom API limits apply and request frequency should be reduced if needed.

Who the server is a good fit for

The server is particularly useful for developer and support teams who want to build their own AI tools on top of Intercom customer data, for example internal research assistants for support, automated reporting on open conversations, or AI-assisted maintenance of help center content. Since Intercom itself has no permanent free tier, only a 14-day trial ahead of its paid plans, the server is best suited to organizations that already run a paying Intercom account and want to weave its data into their own AI workflows — as an alternative to a manual integration against the raw Intercom REST API.

An Intercom account in a US- or EU-hosted workspace (AU is not supported), an MCP client with remote MCP support (or the mcp-remote helper tool), and either OAuth consent or a bearer token with the required permissions.

In MCP clients with native remote support, the region-specific server URL is enough. For clients without native support, via mcp-remote:

{
  "mcpServers": {
    "intercom": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.intercom.com/mcp"]
    }
  }
}

For EU-hosted workspaces, replace https://mcp.intercom.com/mcp with https://mcp.eu.intercom.com/mcp. On first connection, the OAuth consent screen opens in the browser; alternatively, a bearer token can be set directly in the client configuration.

npx mcp-remote https://mcp.intercom.com/mcp

OAuth with a browser-based consent screen (recommended by Intercom) or a direct bearer token stored as credentials in the client configuration.

Access scope depends on the scopes chosen during OAuth consent or the bearer token used: documented scopes are "Read conversations," "Write conversations," "Read and list users and companies," and "Read and write articles."

Requests go through Intercom-hosted endpoints running on Cloudflare infrastructure — mcp.intercom.com (US) or mcp.eu.intercom.com (EU); per the documentation, requests to the EU endpoint are processed within the EU. Responses can include conversation, contact, company, and help center data from the connected Intercom account and are returned to the MCP client and its language model.

The server offers write access to help center articles and internal team notes; customer conversation content also flows into model requests. Bearer tokens should be stored securely like any other API key, and OAuth scopes should be restricted to the minimum actually needed. Manipulated customer or article content could theoretically carry prompt injection attempts against the connected AI model; review agent outputs before any production write operations.