Intercom MCP Server

An MCP (Model Context Protocol) server that connects Cursor to Intercom. Use it from Cursor to list conversations, read threads, draft and send replies, and pull in your support response templates (Do remember to include your own response.md and change the rules as you see fit)

Tools

Each tool is exposed to Cursor (and any MCP client) when this server is running. Prefer draft_reply for AI-generated text so a human can review before the customer sees anything.

Read-only tools

• search_conversations — Search and filter Intercom conversations by state (open, closed, snoozed), date range (created_after, created_before, updated_after, updated_before), and source type (email, chat, push, etc.). Returns a paginated list with conversation IDs, titles, previews, and metadata. Use this to find conversations that match criteria (e.g. “open conversations updated in the last 7 days”).

• get_conversation — Fetch one conversation by ID. Returns the full message thread (who said what, when), contact info, and conversation stats (e.g. time to first reply, reopen count). Use this to read the full context before drafting or sending a reply.

• get_unreplied_conversations — Returns open conversations where the last message is from the customer (i.e. no agent reply yet). Helps prioritize which conversations need a response. You can cap how many open conversations are fetched per page.

• get_new_messages — Poll for new or updated conversations since the last time you checked. Uses a high-water mark: the first call returns conversations from the last hour; later calls return only newer activity. Optional reset flag resets the mark to “now”. Useful for “what’s new since I last looked?” workflows.

Draft (internal note)

• draft_reply — Adds an internal admin note to a conversation. The customer does not see it. The note is signed with - cursor MCP at the end. Use this for AI-generated reply text: the agent reviews and edits in Intercom, then either sends manually from Intercom or uses send_reply with the final text. Safe to allow the model to call.

Send reply (customer-visible)

• send_reply — Sends a customer-visible reply to a conversation. The message is posted immediately as an agent reply; the customer sees it in the conversation.

  Do not add send_reply to the allowlist unless you are extremely sure the model is correct.
  If the model can call send_reply without human approval, it can send wrong, off-brand, or inappropriate messages directly to customers. Prefer keeping send_reply off the allowlist and having humans send the final reply from Intercom (or approve each use). Use draft_reply for AI output and only send after review.

get_new_messages — how it works

The tool keeps an in-memory lastCheckTimestamp. On the first call (when it's null), it uses now - 3600 (last hour) as the since value. On subsequent calls, it uses lastCheckTimestamp as since, so only conversations updated after that time are returned. After each successful fetch, it updates lastCheckTimestamp to now. If you pass reset: true, it sets the mark to now and returns nothing—useful to start fresh. The Intercom API is called with updated_at > since and returns up to 50 conversations.

Note: The high-water mark is in-memory only; restarting the MCP server resets it, so the next call will again use "last hour".

Prerequisites

• Node.js (v18+)
• Intercom API access (API token and an admin/agent ID for sending replies)

Setup

1. Install dependencies

npm install

2. Environment variables

Copy the example env file and set your Intercom credentials:

cp .env.example .env

Edit .env:

Variable	Description
INTERCOM_API_KEY	Your Intercom API token (required).
INTERCOM_ADMIN_ID	Intercom admin/agent ID used as the author of replies and notes (required).
TEMPLATE_DOC_PATH	Path to your response templates file (default: ./templates/responses.md).

3. Build

npm run build

4. Configure Cursor to use this MCP server

Edit .cursor/mcp.json so the server runs from this repo (use an absolute path to dist/index.js if needed) and that env includes your real INTERCOM_API_KEY and INTERCOM_ADMIN_ID. Example:

{
  "mcpServers": {
    "intercom": {
      "command": "node",
      "args": ["/path/to/Cursor MCP/dist/index.js"],
    }
  }
}

Restart Cursor (or reload the MCP server) so it picks up the config and the built server.

Usage in Cursor

Once the MCP server is running in Cursor, you can ask the AI to:

• “Fetch all unreplied Intercom conversations and draft replies.”
• “Get conversation 123456 and suggest a reply using the templates.”
• “Search open conversations updated in the last 7 days.”
• “Draft a reply to conversation 123456 with: [your message].”

Drafts created with draft_reply appear in Intercom as internal notes and end with - cursor MCP. Review and edit them in Intercom, then send the final reply from Intercom (or via send_reply if you have it enabled). Do not add send_reply to the allowlist unless you are extremely sure the model is correct — use draft-first and human review to avoid sending wrong or off-brand messages to customers.

Response templates

Keep your support playbooks in templates/responses.md: agent persona, vocabulary rules, and workflows for common cases (deposits, payouts, KYC, bugs, etc.). The project includes a Cursor rule that tells the AI to check this file when drafting Intercom replies.

Project structure

├── src/
│   ├── index.ts              # MCP server and tool registration
│   ├── lib/
│   │   └── intercom-client.ts # Intercom API client
│   └── tools/
│       ├── draft-reply.ts    # Internal note (draft)
│       ├── send-reply.ts     # Customer-visible reply
│       ├── get-conversation.ts
│       ├── get-unreplied-conversations.ts
│       ├── get-new-messages.ts
│       └── search-conversations.ts
├── templates/
│   └── responses.md          # Support templates (persona, workflows)
├── .env.example
├── .cursor/
│   └── mcp.example.json      # Example Cursor MCP config
└── package.json

Scripts

Command	Description
npm run build	Compile TypeScript to dist/.
npm start	Run the server (node dist/index.js).
npm run dev	Run tsc --watch for development.

License

Private / internal use. Adjust as needed for your setup.