HubSpot Auditor

Read-only audit of a HubSpot portal. You get a scored report of what works, what is broken, what is missing, and what still needs a person to confirm, plus a 30/60/90-day fix plan.

The server speaks MCP over stdio. A CLI runs the same audit. It never sends PUT, PATCH, or DELETE to HubSpot. Its only POST is the start-up read of the token's own scopes.

What's new in v1.5.2: a start-up check warns when the token has a write scope and every HubSpot request is held to a read-only allowlist, a new check reports the original source on closed-won deals, checks no longer pass on partly loaded or unsampled data, and closed stages are left out of the stage-probability check. Thanks to @harrisonoconnorhover for four of those fixes. See CHANGELOG.md.

What's new in v1.5.1: the unused custom properties check no longer passes, or suggests archiving, when it has no sampled values to go on. Thanks to @harrisonoconnorhover. See CHANGELOG.md.

What's new in v1.5.0: findings are ranked by the business process they affect, each fix names an owner and a retest, and forms without captcha or a honeypot are flagged. See CHANGELOG.md.

What's new in v1.4.5: tagged releases publish to npm with GitHub Trusted Publishing. See CHANGELOG.md.

What's new in v1.4.4: site checks block more non-public addresses, and PDF reports stay inside the reports folder. Please update. See CHANGELOG.md.

What's new in v1.4.2: registry metadata for the official MCP Registry and Smithery. See CHANGELOG.md.

What's new in v1.4.1: the npm package record uses the current account email. See CHANGELOG.md.

What's new in v1.4.0: the CLI and MCP server ship as compiled JavaScript. Install with npx. See CHANGELOG.md.

Quick start

You need Node.js 22 or newer.

npx -y hubspot-auditor --demo --out reports

Open reports/example-report.md.

Related MCP server: hubspot-mcp

Who is this for?

RevOps lead cleaning a portal before a process change. You need one pass over duplicates, pipelines, and forms.

• "Audit this portal in demo mode and show the broken checks."

• "Which required contact properties are thin?"

• "Which deal stages have a weak probability?"

You get findings tagged with the business process they affect, plus severity, evidence, and a recommended fix.

Administrator checking consent and sending. You need GDPR fields, spam controls, and sender authentication called out first.

• "Run the demo audit and list the critical findings."

• "What does the email auditor check?"

• "Show the 30-day plan."

You get each fix with a suggested owner role and a retest. Billing and lead routing come before a long list of lower-impact items.

Consultant handing a client a checklist. You need a report you can regenerate after they change a YAML threshold.

• "List every check and its severity."

• "Run the fictional Harbour Glass audit."

• "Write the Markdown and JSON reports."

You get examples/example-report.md and examples/example-report.json from the sample portal. No customer data is in that example.

Agent user who already has HubSpot connected. You want a second server that only scores the portal.

• "Run a read-only audit. Do not change the portal."

• "Use demo true if no token is set."

You get the same report. The token stays in the environment. It is not a tool argument.

What it checks

Rules live in versioned YAML under rules/. Each file has version: 1 and a list of checks. Turn a check off with enabled: false. Thresholds sit under params.

A check returns working, broken, missing, or needs_operator_confirmation. Evidence is counts and property names. The report does not print contact records.

Spam protection lists each live form that has no captcha and no honeypot field. A form on a public page with recent submissions ranks above a form that is not embedded, even when the private form has more submissions. The count comes from a read of the last 30 days when the submissions API allows it. When that read is missing, the report says the count was not returned. When captcha and honeypot cannot be confirmed from the form definition, the check is needs_operator_confirmation rather than a pass. The forms API exposes recaptchaEnabled. It does not return HubSpot's honeypot toggle, so a hidden honeypot field is the field signal, and the finding asks you to confirm the toggle in the form editor.

Findings are ordered by the business process they affect, then by severity. A single territory or billing fault outranks a longer list of unused fields. The count in the evidence does not set the rank.

The score is the percentage of checks that are working. A check that needs operator confirmation is not a pass. The 30-day plan is critical and high broken items. The 60-day plan is medium broken items, plus critical and high missing items, plus critical and high checks that need confirmation. The 90-day plan is what remains. Each plan item names an owner role and a retest.

These checks stay at needs_operator_confirmation on a pass until you set params.operatorConfirmed: true for that check: required properties, lifecycle and lead status, stage probabilities, stale deals, required stage fields, and stage automation. The flag means you accept the rule in the YAML. It does not change the portal. A broken or missing result is unchanged.

Closed-won deal attribution counts a deal when hs_is_closed_won is the string true. When that property is absent, a closed stage at probability 100 counts, and so does probability 100 when the stage has no closed flag. Fewer than five sampled closed-won deals is missing. An original-source share below 0.80 is broken. Where the campaign property is present, the count of deals with a value is evidence only and does not change the status. Ad spend is always needs_operator_confirmation. Evidence states the sample size and leaves out deal names, amounts, contacts, and ids.

Live mode reads the private app token's scopes at start-up, then reads HubSpot with allowlisted GET requests. If a call is forbidden or the payload has no field for that check, the finding is missing. A failed ticket-pipeline read does not hide the deal pipeline. Fill rates count only contact properties that were requested. Demo mode loads demo/portal.json, a fictional company called Harbour Glass Co.

Logo, favicon, heading, and colour checks run on the fetched landing page. Add params.markers in rules/brand.yaml only when a phrase must appear verbatim. The buyer-stage patterns live in rules/journey.yaml. Pass --site to judge a specific public page. If you omit it, tracking uses the portal's primary site domain. Landing page checks use the CMS page URL.

MCP tools

Example workflow

You: "Run the Harbour Glass demo and tell me the score."

The server calls run_audit with demo: true.

You get a score of 33, with 16 working, 21 broken, 11 missing, and 0 that need operator confirmation. The first 30-day items are required properties, workflow health, and form field mapping. Spam protection is also in that plan, after the higher-impact faults.

Setup

You need Node.js 22 or newer.

npm install -g hubspot-auditor
hubspot-auditor --demo --out reports

Or, from a clone:

git clone https://github.com/M-Pineapple/hubspot-auditor.git
cd hubspot-auditor
npm install
npm run build
npm test
npm run audit:demo

npm run audit:demo writes examples/example-report.md, examples/example-report.json, and examples/example-report.pdf.

For a live portal, create a HubSpot private app and copy a token that has read scopes only. Do not grant write scopes. Put the token in the environment. Do not pass it as a tool argument.

At start-up the auditor reads that token's scopes with one POST to /oauth/v2/private-apps/get/access-token-info. It warns when a write scope is present. It never sends any other non-GET request.

export HUBSPOT_ACCESS_TOKEN="paste-a-read-only-private-app-token"
hubspot-auditor --site https://example.com --summary

--site is the public page the tracking checks fetch. --summary prints counts and check ids only.

No write scope is required. A missing read scope becomes a missing finding for that area. HTTP 429 and 5xx are reported as throttling or an outage, not as a missing scope.

Sampling

Each of contacts, companies, and deals is read up to 500 records (5 pages of 100). Form detail is read for up to 30 forms. Published pages are read for up to 20 pages for tracking. Campaigns are one page of up to 100. Ad-campaign assets are read for at most 20 of those campaigns. The results describe that sample, not the whole portal.

The audit does not call webhook subscription edits or a connected-app list. Those HubSpot URLs do not return that data for a private app.

If the system resolver cannot find api.hubapi.com, API reads fall back to DNS over HTTPS at 1.1.1.1 and still check the original host name on the certificate. Sender TXT lookups use the system resolver first. Set HUBSPOT_AUDITOR_DOH=0 to turn the fallback off. The fallback is on unless you set that variable.

publicSiteUrl must be https. Hosts that resolve to loopback, private, link-local, or CGNAT addresses are refused, and each redirect is checked again.

MCP config

Paste this into Claude Desktop, Cursor, or Claude Code. Replace the token. For a demo-only server, you can omit the env block and call run_audit with demo: true.

{
  "mcpServers": {
    "hubspot-auditor": {
      "command": "npx",
      "args": ["-y", "hubspot-auditor"],
      "env": {
        "HUBSPOT_ACCESS_TOKEN": "paste-a-read-only-private-app-token"
      }
    }
  }
}

The same file lives in config/claude-desktop.json, config/cursor-mcp.json, and config/claude-code.json.

From a git clone, run npm run build and point the command at node with args set to ["/path/to/hubspot-auditor/dist/main.js"].

Start the server on its own with:

npm start

How this differs from the official HubSpot MCP

Use the official server when you want an agent to work in the CRM. Use this server when you want a read-only audit.

Example report

The full fictional report is examples/example-report.md. The JSON form is examples/example-report.json. The PDF is examples/example-report.pdf.

The checked-in sample scores 33 (16 working, 21 broken, 11 missing, and none need operator confirmation). The token-scope check passes, and closed-won attribution passes on the sample. The public demo form has no captcha and 12 submissions in 30 days, so spam protection is critical. On this sample the passes that would need confirmation are already broken or missing.

# HubSpot audit: Harbour Glass Co

This report uses a fictional sample portal. It is not a customer account.

Score: **33** (share of checks that are working). A check that needs operator confirmation is not a pass.

| Status | Count |
| --- | ---: |
| Working | 16 |
| Broken | 21 |
| Missing | 11 |
| Needs operator confirmation | 0 |

The broken list starts with required properties, workflow health, and form field mapping. Spam protection is critical and sits later in that list, because conversion ranks after lead routing. The 60-day plan starts with the missing deal amount, because billing outranks the medium items in that bucket.

Extend a check

See CONTRIBUTING.md.

More MCP servers by M-Pineapple

• Dev Project Coordinator

• CryptoAnalysisMCP

• warp-command-runner

• msty-admin-mcp

• member-berries-apple-mcp

• seashell

• TextEdit-MCP

💖 Support This Project

If HubSpot Auditor has helped you audit a portal, consider supporting its development:

Licence

MIT. Copyright (c) 2026 M-Pineapple.