HubSpot MCP Server

A Model Context Protocol (MCP) server that exposes HubSpot CRM data and actions as tools for AI agents (Claude, Cursor, and any other MCP-compatible client) — so an agent can look up a contact, search companies, upsert a lead, and log a call or note, all through typed, validated tool calls instead of hand-rolled API glue.

Battle-tested against a live HubSpot account: contact lookup/upsert, company search, and note logging all verified end-to-end (see Tools for sample I/O).

If you find this useful, a star helps other people building MCP integrations find it.

Architecture

src/
  hubspotClient.ts   HubSpot REST API wrapper (auth, requests, 429 retry, typed responses)
  index.ts           MCP server: tool definitions + stdio transport wiring

• Transport: StdioServerTransport — the server communicates with its MCP client over stdin/stdout, so it's launched as a subprocess (no network port to manage).
• Auth: A HubSpot Private App access token is read from the HUBSPOT_ACCESS_TOKEN environment variable at startup. The token is never hardcoded or logged.
• Error handling: All HubSpot API calls go through HubSpotClient, which:
  • Converts network/HTTP failures into a typed HubSpotApiError with a client-safe message.
  • Retries 429 Too Many Requests responses (respecting the Retry-After header, with exponential backoff as a fallback) up to 2 times before surfacing the error.
  • Never throws out of a tool handler — index.ts catches every error and returns an MCP isError: true tool result instead of crashing the server.
• Tool schemas: Each tool's input is defined with zod, giving strict runtime validation and auto-generated JSON Schema for the MCP client.

Setup

cd hubspot-mcp-server
npm install
cp .env.example .env   # then fill in HUBSPOT_ACCESS_TOKEN
npm run build

Generate a HubSpot Private App access token: HubSpot → Settings → Integrations → Private Apps → Create a private app. Grant at minimum:

• crm.objects.contacts.read / crm.objects.contacts.write
• crm.objects.companies.read
• crm.objects.notes.write (or the relevant engagements scope for the activity types you plan to log)

Running standalone

HUBSPOT_ACCESS_TOKEN=pat-xxxxx npm start

For local development without a build step:

HUBSPOT_ACCESS_TOKEN=pat-xxxxx npm run dev

Configuring Claude Desktop

Add this server to your claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "hubspot": {
      "command": "node",
      "args": ["/absolute/path/to/hubspot-mcp-server/dist/index.js"],
      "env": {
        "HUBSPOT_ACCESS_TOKEN": "pat-xxxxx"
      }
    }
  }
}

Restart Claude Desktop after saving. The four HubSpot tools will appear in the tool picker.

Tools

hubspot_get_contact

Look up a contact by email.

{ "email": "jane.doe@example.com" }

Returns first name, last name, lifecycle stage, and associated company name (if any).

hubspot_search_companies

Search companies by name or domain.

{ "query": "acme" }

Returns matching company names and domains.

hubspot_create_contact

Create a contact, or update it if the email already exists.

{
  "email": "jane.doe@example.com",
  "firstName": "Jane",
  "lastName": "Doe",
  "company": "Acme Corp"
}

hubspot_log_activity

Log an engagement on a contact's timeline.

{
  "contactId": "12345",
  "activityType": "NOTE",
  "body": "Discussed renewal timeline on the call."
}

activityType is one of NOTE, EMAIL, or CALL.

Notes

• Rate limits: HubSpot returns 429 when the account's rate limit is exceeded. The client retries automatically; if retries are exhausted, the tool returns a clear error message rather than crashing.
• All tool errors (invalid input, HubSpot API errors, network failures) are returned as MCP tool errors (isError: true) so the calling agent can react instead of the server process dying.

License

MIT — see LICENSE.