Connecting AI Assistants (MCP)

Turn on your bot's MCP channel to let ChatGPT and other AI assistants search and read its knowledge base. Covers setup, what an assistant can reach, privacy, and troubleshooting.

The MCP channel makes your bot's knowledge base searchable from inside an AI assistant such as ChatGPT. Once connected, you and your team can ask the assistant questions about your own documentation and get answers that link back to your pages. This article covers turning the channel on, adding it to ChatGPT, what an assistant can and cannot reach, and what to check when it doesn't work.

What MCP is

MCP (Model Context Protocol) is a standard way for AI assistants to connect to outside sources of information. Think of it as a plug socket: any assistant that supports MCP can plug in to your bot.

Turning the channel on gives your bot its own connection URL. An assistant you add that URL to gets two read-only abilities:

Search: find documents in the bot's knowledge base that match a question.
Read: get the full text of a document that search found.

Typical uses:

A support agent drafting a reply who needs the exact wording of a policy.
Someone in sales checking what the docs say before answering a prospect.
Anyone on the team asking "where is this documented?" without hunting through the help center.

Questions asked through this channel do not use your monthly interactions, and the channel is available on every plan.

Before you start

A bot that has finished training. A bot with no processed sources returns nothing.
An AI assistant account you can change settings on. For ChatGPT, that means a paid ChatGPT plan.
A few minutes to read the Privacy section further down this page. The connection has no password.

Turn the channel on

The MCP channel is off by default for every bot.

Open your bot and click Deploy.
Open the MCP tab.
Click Enable MCP (or flip the toggle at the top right of the tab).
Click Save settings. The badge next to the toggle shows Active.

The channel is live only after you save. Until then the tab shows "You have unsaved changes. The connection stays off until you save."

Copy your connection URL

With the channel on, the MCP tab shows a Connection URL with a Copy URL button. It looks like this:

https://app.helpjet.com/api/mcp/bot/YOUR_BOT_ID

Copy it from the tab instead of typing it. Each bot has its own URL, so connecting three bots means adding three connections.

Add it to ChatGPT

This is a one-time setup on chatgpt.com in a browser, and takes about two minutes. The same steps are shown on the MCP tab under Add to ChatGPT.

In ChatGPT, open Settings → Security and login and turn on Developer mode. This needs a paid ChatGPT plan.
Go to chatgpt.com/plugins, click +, and choose Create custom MCP server. (Create plugin and Upload plugin are for something else.)
Give the connection a name you will recognize, such as your bot's name, and paste your connection URL into the server URL field.
Set authentication to No authentication. There is no password or API key to enter.
Confirm that you trust the server if asked, then create it. ChatGPT should list two tools: search and fetch.
Start a new chat, click + in the message box, choose Developer mode, and select the connection you created.

Check it works

Ask something your documentation definitely covers, and ask for sources. For example:

Search my HelpJet knowledge base for our refund policy and cite the sources.

A good answer uses your own wording and links to your own pages. If the answer is general and quotes nothing of yours, the assistant probably didn't use the connection; mention the connection by name in your question and try again.

Missing links are not a sign of a problem on their own. Uploaded files, pasted text, and Q&A entries have no public URL, so the assistant can only refer to them by title.

Other AI assistants

Any assistant that can connect to a remote MCP server by URL, without a password, can use the same connection URL. Look for "custom connector" or "MCP server" in its settings.

ChatGPT is the only assistant with written steps so far. Step-by-step instructions for other assistants will be added as each one is verified.

What an assistant can reach

An assistant can search and read in full:

Website pages, crawls, and sitemaps
WordPress content
PDFs and uploaded files
Pasted text
YouTube transcripts
Q&A entries saved in Train & Test

An assistant can never reach:

Anything ingested from Help Scout conversations
Your other bots, or anyone else's
Visitor conversations, analytics, billing, or team details
Any way to add, edit, or delete content or settings

A few details worth knowing:

The connection is read-only. An assistant can look things up; it can never change anything.
Long documents are cut at 50,000 characters. The assistant is told when a document was cut. If a long manual matters, split it into several training sources.
Sources being refreshed are briefly unavailable. A source waiting to be re-trained is left out of search and read until the refresh finishes.
Requests are not shown in your dashboard. Questions asked through an assistant do not appear in the Activity tab or in your bot's metrics.

Privacy: read before turning it on

It is the same exposure as your public chat widget, with one difference. The widget answers one question at a time; this channel hands over whole documents on request. Share the URL with that in mind.
Your content is sent to the assistant's provider. Whenever a connected assistant searches or reads a document, the matching text leaves HelpJet and goes to that provider (OpenAI, for ChatGPT). How they handle it is governed by your account with them.
Help Scout conversations are always excluded. They contain your customers' own words, so they are never served on this channel. You can't switch this exclusion off, and your widget still answers from them as before.
That is the only exclusion. An internal PDF, a crawled staging site, or a pasted internal runbook is fully reachable.

Leave the channel off for any bot trained on material that is not already public. If you need an internal-only bot, keep it as a separate bot with MCP off.

Turn the channel off

Open Deploy, switch to the MCP tab, flip the toggle off, and click Save settings.

This cuts every connection to that bot immediately, in every assistant, including connections other people added. Removing the connection inside an assistant only affects that one account.

Turning the channel off and on again does not change the URL. If the URL has been shared more widely than you intended, the way to stop access is to leave the channel off.

Troubleshooting

Every refusal looks the same from inside an assistant: the connection adds but lists no tools, or every question comes back with nothing. That is deliberate (the connection never reveals whether a bot exists), so work through the list.

Troubleshooting
Symptom	What to check
Won't connect, or connects with no tools	The MCP tab shows Active and you clicked Save settings. The URL was copied from the tab, with no trailing slash or spaces. The bot itself is Active; an Inactive, Archived, or Building bot refuses every request.
"Developer mode" isn't in ChatGPT settings	Look under Settings → Security and login. If it isn't there, your ChatGPT plan or workspace doesn't have it; a workspace admin may need to enable it.
Searches come back empty	The bot's sources show Processed, not pending or failed. Try the wording your documentation itself uses. A bot trained mostly on Help Scout conversations has little for this channel to search.
Answers are out of date	Refresh the training source on the Sources tab. See Training Sources Overview.
No citations, or an invented link	Ask the assistant to cite its sources. A link to a page that isn't yours means it answered from general knowledge, not your knowledge base.
"Too many requests"	The bot reached its per-minute request limit. Wait a minute and try again. If it keeps happening with light use, someone else may be using the URL.
A long document is cut off	Expected at 50,000 characters. Split the source into smaller ones.
Answers don't sound like your bot	Expected. The assistant writes in its own voice. Use the Bot Widget, Page Embed, or Chat Link where voice matters.

Common questions

Does this use my interactions?

No. Questions asked through a connected assistant don't count against your allowance. See Understanding Interactions for what does.

Can an assistant change my bot or its training?

No. The connection can only search and read. It can't add or remove sources, edit settings, or delete anything.

Can I protect the URL with a password?

Not currently. Treat the URL like a public link, and leave the channel off for bots trained on private material.

Can I connect more than one bot?

Yes. Each bot has its own URL, and each one is added to your assistant as a separate connection.

Where to go next

Embedding & Sharing Your Bot: the channels that put your bot in front of visitors.
Training Sources Overview: what the assistant searches is what your bot is trained on.
Security & Privacy: how HelpJet handles your data.