ChatGPT integration

Bring your help center into ChatGPT.

Your team can ask ChatGPT what customers could not find, draft the missing articles and keep the rest current, without leaving the chat. Your customers can connect it too, and get answers from your published articles, linked to the source.

For your team on Growth and Catalyst, and in their 14-day trials. For your customers on every plan.

ChatGPT with developer mode, on the web
  • For your team Growth and Catalyst

    The HelpCenter.io MCP server

    Search, write, organize and measure your knowledge base: 29 tools, inside the permissions you approve.

    https://mcp.helpcenter.io
    • Sign in with HelpCenter.io
    • Read and write
    • Asks before it writes
  • For your customers Every plan

    Your help center's own MCP server

    Every live, public help center open to search engines has one, on by default. Customers add it to ChatGPT and get answers from your published articles.

    https://your-help-center/mcp
    • No sign-in
    • Read-only
    • Published articles only
Your team, in ChatGPT

Answer the customer, then fix the article they needed.

Paste a customer's question and ChatGPT checks it against your help center: what you already cover, what readers searched for and missed, and what to change. It drafts the reply with links to your articles, and the fix for the article itself.

Nothing changes until you say so: by default ChatGPT asks before every write, and on Catalyst an edit to a live article can be staged privately for your team to publish.

  1. Read your help center Search and read every article in it.
  2. See what readers missed The searches that found nothing, ranked by demand.
  3. Propose the fix A draft, an edit, or a staged change for review.
Northwind help center Example
  1. You: A customer asks how to change the email they sign in with. Draft a reply from our help center.

  2. Tool call: helpcenter_search_articles reads content

  3. Assistant: “Update your profile” covers it. Here is a reply you can send:

    Go to Profile → Email and enter the new address. We send a confirmation link there, and the change applies once you click it. help.northwind.energy/account/update-your-profile
  4. Tool call: helpcenter_get_search_queries reads analytics

  5. Assistant: One more thing: “change email” was searched 23 times last month. Results came back, but nobody opened them, because the title never mentions email. Shall I fix the title?

  6. You: Yes. Stage it so the team can review it.

  7. Tool call: helpcenter_stage_article_changes confirmed by you, not live yet

  8. Assistant: Staged. Readers keep seeing the current title until your team publishes the change.

    Change your sign-in email or profile Staged
An example on Northwind, a fictional help center. The numbers are illustrative; the tools are the real ones ChatGPT calls. Staging needs Catalyst. On Growth, ChatGPT can update the article directly once you confirm.
29 tools, one conversation

Ask for the work, not the clicks.

Things your team can ask ChatGPT today. Each one runs on tools HelpCenter.io gives it, inside the permissions you approve.

See all 29 tools
  • Find the gaps

    Ask: What did customers search for last month and not find?

    The searches that came back empty, ranked by demand. Plus the ones where results showed and nobody clicked, which usually means a title to fix rather than an article to write.

    • helpcenter_get_search_queries
    Needs: Analytics
  • Fix what readers rate poorly

    Ask: Which of our most-read articles are rated worst this quarter? Suggest fixes for the top three.

    The worst-rated articles first, weighted towards the ones people actually read, and each one read in full before a fix is suggested.

    • helpcenter_get_content_performance
    • helpcenter_get_article
    Needs: Analytics Read content
  • Reply from your docs

    Ask: Draft a reply to this customer email from our help center, and link the articles you used.

    A reply grounded in what you have actually published, with links your customer can follow. Nothing in your help center changes.

    • helpcenter_search_articles
    • helpcenter_get_article
    Needs: Read content
  • Turn a thread into an article

    Ask: Turn this support thread into a help article and save it as a draft in Billing.

    A new article in the category you name, saved as a draft, so publishing stays your call.

    • helpcenter_create_article
    Needs: Write content
  • Translate an article

    Ask: Translate “Refunds on annual plans” into German and add it as the German version.

    The German title and body saved on the same article, in any language your help center has switched on.

    • helpcenter_get_article
    • helpcenter_update_article
    Needs: Write content
  • Tidy your categories

    Ask: Move every article about invoices into a new Invoices category, and write it a one-line description.

    The new category with its description, and each article moved into it. Nothing is deleted.

    • helpcenter_create_category
    • helpcenter_update_article
    Needs: Write content
A Northwind customer's chat Example
  1. Customer: How do I download my invoices from Northwind?

  2. Tool calls: search fetch Northwind's help center, no sign-in

  3. Assistant: Open Billing → Invoices in Northwind, pick the months you need and choose Download. You get one PDF per invoice, or a single CSV for your accountant.

    Source: help.northwind.energy/billing/download-your-invoices

An example on Northwind, a fictional help center. The customer's ChatGPT searched and read its published articles through its public MCP server, and linked the one it used.
Your customers, in their ChatGPT

Answers from your articles, not a guess from the web.

Every live, public HelpCenter.io help center that search engines may index is also an MCP server, at its own address with /mcp on the end. A customer who adds it to ChatGPT gets answers searched and read from your published articles, with a link to the one it used.

search
Finds published articles, best match first, each with its link and a short snippet.
fetch
Reads one article or category in full, as Markdown.
list_categories
Shows how the help center is organised, and how much each category holds.
  • No sign-in and no API key, just as your help center needs none.
  • Only what an anonymous visitor can read. Drafts, private and unlisted articles and your team's notes never leave.
  • It asks ChatGPT to prefer your articles over what it already knows, and to cite the one it used.
  • Its search and fetch take the exact shape ChatGPT's company knowledge and deep research ask for, so they work there too.
  • On by default on every plan. Switch it off in Settings → Public MCP server.
  • Readers can find it under Use with AI in the article menu, with the steps for ChatGPT, Claude, Cursor and VS Code.

Your customers add it with developer mode in ChatGPT on the web, or their workspace admin adds it for everyone. Private, password-protected, IP-restricted and embed-only help centers have no public server, because an anonymous visitor could not read them either. Nor does one that has turned off Allow search engines, or one that is not live.

Connect in four steps

One URL, one sign-in, and ChatGPT is connected.

In ChatGPT on the web:

  1. Turn on developer mode It is a switch in ChatGPT's settings. On Business, Enterprise and Edu workspaces, an admin turns it on first. OpenAI's guide shows where it is on your plan.
  2. Create an app for HelpCenter.io Create a developer-mode app, paste https://mcp.helpcenter.io as its MCP server URL, and pick OAuth for authentication.
  3. Sign in to HelpCenter.io ChatGPT opens our sign-in and consent screen. Pick the help centers it can reach, review the permissions it asks for, and allow.
  4. Ask your first question In a new chat, choose HelpCenter.io from the + menu or @mention it, then ask what your customers searched for last month and could not find.

Customers connect your public server the same way, with your help center's /mcp URL and No Authentication.

For your whole ChatGPT workspace

On Business, Enterprise and Edu, an admin creates the app once and publishes it from Workspace settings → Apps. Teammates then find it in their own Apps settings and sign in to HelpCenter.io as themselves, so each person reaches only the help centers they already work on.

On Enterprise and Edu, admins can also choose which of its actions the workspace may use before they publish it.

OpenAI Responses API
{
  "type": "mcp",
  "server_label": "helpcenter",
  "server_url": "https://mcp.helpcenter.io",
  "authorization": "<your HelpCenter.io API key>",
  "allowed_tools": [
    "helpcenter_search_articles",
    "helpcenter_get_article"
  ],
  "require_approval": "never"
}

One entry in the tools array: an agent that searches and reads your articles, and can change nothing, so it needs no approval prompts. Drop allowed_tools to offer all 29, and keep approval on for the ones that write. Keys come from your help center's Settings → API Keys.

Codex
codex mcp add helpcenter --url https://mcp.helpcenter.io
codex mcp login helpcenter

The second command opens the HelpCenter.io sign-in in your browser.

Security and control

You decide how far it goes.

Access is granted help center by help center, and you can take it back whenever you like. Your password never reaches ChatGPT: it signs in on our own consent screen.

  • Your consent screen, not a pasted key

    ChatGPT signs in through HelpCenter.io over OAuth 2.1. You choose which help centers it can reach and see every permission it asks for before you allow it.

  • ChatGPT asks before it writes

    Every HelpCenter.io tool says whether it only reads, and the destructive ones say so too. ChatGPT treats anything that is not read-only as a write, and asks you before running it unless you change its approval settings.

  • Drafts, and staged edits on Catalyst

    New articles can be saved as drafts, so publishing stays your call. On Catalyst, an edit to a live article can be staged privately for your team to publish or discard.

  • Change your mind any time

    Change which help centers ChatGPT can reach, or disconnect it, in Account → Connected Apps. The next request it makes is refused.

  • Reader privacy by default

    Reader email addresses are withheld unless a connection with write access asks for them. IP addresses are never exposed, and search data is aggregated daily.

  • The public server shows only what is public

    A help center's own MCP server answers as an anonymous visitor would be answered: published, public articles. Drafts, private and unlisted articles and your team's notes never leave.

Availability

Your customers on every plan. Your team from Growth.

No add-on and no charge per tool call. The public server comes with every help center, and the MCP server for your team comes with the plan, 14-day trial included.

Every plan

For your customers

from $29 /mo, billed annually

Bootstrap, or $39 month to month

  • Your help center's own public MCP server, on by default
  • search, fetch and list_categories, read-only
  • Published, public articles only
Start free trial

Growth

For your team

$119 /mo, billed annually

or $149 month to month

  • ChatGPT for your team: articles, categories, comments, editorial notes and analytics
  • Up to 10 teammates
  • Not included: Staged changes and version history, which need Catalyst
Start a Growth trial

Catalyst

All 29 tools

$179 /mo, billed annually

or $199 month to month

  • Every one of the 29 tools, staged changes and version history included
  • Unlimited teammates
  • AI question answering and translations in your help center too
Start a Catalyst trial

On the ChatGPT side, adding an app by URL takes developer mode, on the web. OpenAI lists full read-and-write support for Business, Enterprise and Edu, and read-only connections on Pro. Your ChatGPT subscription is billed by OpenAI. OpenAI's current requirements.

Bootstrap and older plans priced below Growth do not include the MCP server for your team. Compare every plan.

FAQ

Questions, answered.

Want the whole tool list? It is on the MCP server page. Something else? Our team answers from our help center.

Is there a HelpCenter.io plugin for ChatGPT?

Not in ChatGPT's Plugin Directory. You connect HelpCenter.io yourself, as a developer-mode app with the URL https://mcp.helpcenter.io, and the connection steps walk through it. The original ChatGPT plugins beta, which this page once took sign-ups for, closed in April 2024. Plugins in ChatGPT today are built on MCP, the open standard both HelpCenter.io servers speak.

Which ChatGPT plans can connect to HelpCenter.io?

Adding a server by URL takes ChatGPT's developer mode, on the web. OpenAI's help center lists full read-and-write support for Business, Enterprise and Edu workspaces, where an admin switches developer mode on, and read-only connections on Pro. It is not available in ChatGPT's mobile apps. OpenAI changes these rules, and its article on developer mode has the current list.

Which HelpCenter.io plans include it?

Connecting ChatGPT for your team is included in the Growth and Catalyst plans, and in the 14-day trials of both. Staged changes and version history need Catalyst. Your help center's own public MCP server, the one your customers connect, is on every plan and on by default.

Does it cost extra?

Not on the HelpCenter.io side. There is no add-on to buy and no charge per tool call, and standard API rate limits apply. Your ChatGPT subscription is separate, and billed by OpenAI.

Can ChatGPT change my help center without asking?

Not by default. ChatGPT treats every tool that is not marked read-only as a write, and asks you before running it unless you change its approval settings. New articles can be saved as drafts, and on Catalyst an edit to a live article can be staged for your team to publish or discard. Deleting a category, discarding a staged edit and restoring an old version are marked destructive.

What can my customers see through our public MCP server?

Exactly what an anonymous visitor to your help center can read: published, public articles and the categories they sit in. Drafts, private and unlisted articles, and your team's internal notes never leave. It is read-only, and you can switch it off in Settings → Public MCP server.

What about custom GPTs and GPT Actions?

OpenAI is retiring custom GPTs on December 11, 2026, and their actions do not carry over: its migration guide says to rebuild them on a connector or a custom MCP server. HelpCenter.io already runs one, so there is no OpenAPI schema to maintain: connect ChatGPT to https://mcp.helpcenter.io, and let your customers use your help center's own /mcp address.

Can I use it with the OpenAI API or Codex?

Yes. Point the Responses API's MCP tool at https://mcp.helpcenter.io with a HelpCenter.io API key, or add the same URL to Codex and sign in. Either way you get the same 29 tools. ChatGPT itself cannot send an API key, which is why it signs in instead.

How do I disconnect ChatGPT?

Open Account → Connected Apps in HelpCenter.io and disconnect it, or change which help centers it can reach. You can also remove the app from ChatGPT's own settings.

Make your help center the source ChatGPT answers from.

Start a 14-day trial, connect ChatGPT, and ask what your customers could not find. No credit card needed.