Building Custom AI Chatbots With ChatGPT and Claude

Created by

Do not index

Created time

May 28, 2026 11:03 AM

Building a custom AI chatbot with ChatGPT and Claude takes two parts: a chatbot trained on your own content that answers visitors with an OpenAI model, and an MCP connection that lets Claude manage that chatbot for you.

TL;DR

There are four ways to build a custom AI chatbot with ChatGPT or Claude: a custom GPT, a Claude project, a coded build on the OpenAI or Anthropic API, or a no-code platform trained on your content.

A custom GPT and a Claude project both live inside the assistant. A chatbot for website visitors needs the API or a platform.

This guide builds the chatbot in SiteGPT with no code. It answers with GPT-4.1 mini or GPT-4.1 and your own pages.

Claude then connects to the remote MCP server at https://sitegpt.ai/mcp. You approve access in the browser, and no API token is pasted into the chat.

MCP, CLI, and API access need the Growth plan or above: $129 a month billed monthly, or $79 a month billed yearly. The 7-day trial requires a credit card.

Key takeaways

Stage	What you do	Where
1. Build	Create a chatbot and add your website with Scrape Website or a sitemap	SiteGPT dashboard
2. Test and install	Ask three or four real questions, then paste the JavaScript embed into your site	Dashboard and your website
3. Connect	Add SiteGPT under Settings, then Connectors, and approve access	Claude and the SiteGPT approval page
4. Manage	Ask Claude to list chatbots, find unanswered questions, add content, or review leads	Claude
Optional	Let an agent build a preview chatbot from a URL before you have an account	SiteGPT CLI

Summary graphic of the four stages: build a chatbot in SiteGPT, install the embed, connect Claude over MCP, and manage the chatbot from Claude

Four ways to build a custom AI chatbot with ChatGPT or Claude

A custom AI chatbot can be built four ways with ChatGPT or Claude, and the right one depends on who will talk to it. Two options stay inside the assistant. Two can answer visitors on a website.

Option	Where the chatbot lives	Code needed	Main limit
Custom GPT in ChatGPT	Inside ChatGPT, through the GPT Store or a shared link	No	Personal ChatGPT accounts cannot create new GPTs, and OpenAI plans to retire custom GPTs in favor of Plugins
Claude project	Inside Claude, as a workspace with its own knowledge base and instructions	No	It serves people who use Claude. Free accounts are limited to five projects
Build on the OpenAI or Anthropic API	Anywhere you host it	Yes	You write and maintain the retrieval, the chat interface, and the hosting
Content-trained platform such as SiteGPT	A widget on your website	No	Paid plans with monthly message quotas, and no free plan

OpenAI's help center says new GPT creation is closed on Free, Go, Plus, and Pro accounts. Business, Enterprise, and Edu workspaces can still create GPTs. The same page lists a planned retirement date of December 11, 2026 for affected Enterprise workspaces.

Anthropic describes projects as self-contained workspaces with their own chat histories and knowledge bases. A project suits a team that asks questions of its own documents inside Claude.

The rest of this guide follows the fourth option. The chatbot answers website visitors, and Claude manages it over MCP. Teams that want the coded route can compare it with how to train ChatGPT on your own data.

What managing a chatbot from Claude means

Managing a chatbot from Claude means Claude calls the chatbot platform's tools for you. The chatbot still runs on your website. Claude works on the account behind it: the content, the settings, the conversations, and the leads.

The connection uses the Model Context Protocol. The MCP documentation defines it as an open-source standard for connecting AI applications to external systems. Anthropic open-sourced MCP on November 25, 2024.

A product exposes an MCP server, and an assistant connects to it as a client. SiteGPT runs a remote MCP server at https://sitegpt.ai/mcp. The concept is covered in more depth in MCP for customer service AI.

The remote server has three tools that reach all of API v2. It also has tools for common jobs such as conversations, leads, analytics, content, and widget appearance.

Tool	What it does	Changes data
`search`	Searches the API v2 specification to find an operation. It makes no API calls	No
`execute_read`	Calls API v2 operations with GET only	No
`execute_write`	Calls operations that create, update, or delete	Yes. It is marked destructive, so the assistant confirms first

What you need before you start

You need three things before you start: a SiteGPT account on the right plan, a website, and Claude.

A SiteGPT plan with API access. The docs list API v2, the CLI, and MCP on Growth and above. Growth costs $129 a month billed monthly, or $79 a month billed yearly ($948 a year). It includes 2 chatbots, 10,000 pages, and 10k GPT-4.1 mini messages a month.

A website address and a way to add a script to the site.

A Claude account. Anthropic's help center lists custom connectors on Free, Pro, Max, Team, and Enterprise plans. Free users get one custom connector.

Starter costs $59 a month billed monthly, or $39 a month billed yearly. It builds and installs the chatbot but does not include API access. On Starter, the approval page shows "API access needs the Growth plan or above."

How to build the chatbot

Building the chatbot takes five steps in the SiteGPT dashboard, and none of them needs code. The steps below follow the SiteGPT quickstart.

Step 1: Sign in and choose a plan

Go to the SiteGPT login page and enter your email address. SiteGPT emails a sign-in link, and there is no password. The link creates an account if you do not have one.

You need an active plan to train a chatbot. Every self-serve plan starts with a 7-day free trial. A credit card is required at checkout.

Step 2: Create a chatbot

On the Chatbots page, select Create New Chatbot. Enter a name your team will recognize, then select Create Chatbot.

Pick a name that reads well in a sentence, such as "Support Bot". You will use that name in prompts to Claude later.

Step 3: Add your website

In the menu, select Website Links, then Scrape Website. Enter your home page address, a recursion depth of 2, and a maximum of 50 pages. Select Add Links.

If your site has a sitemap, use Add from Sitemap instead. You can also upload files, add text snippets, add YouTube videos, or connect Notion, Google Drive, Dropbox, OneDrive, Box, or GitHub.

SiteGPT reads and indexes each page in the background. The Links List shows how many pages are Trained, Pending, and Failed. The chatbot answers by retrieval, which what is RAG explains. For what to include and what to leave out, see how to train a chatbot on your own data.

Step 4: Test the chatbot

In the menu, select Dashboard, then Chat with your chatbot. Ask three or four questions your visitors often ask, such as prices, opening hours, or how to contact you.

Each answer comes from your website. If you ask about something that is not on the site, the chatbot usually replies that it does not have enough information.

Step 5: Add the chatbot to your website

In the menu, select Installation. Under JavaScript Embed, copy the code and paste it into the head section of every page. Publish the change.

The docs have separate steps for WordPress, Shopify, Wix, Squarespace, and Webflow. Open your site in a new browser window to check that the chat icon appears.

How to connect the chatbot to Claude over MCP

Connecting the chatbot to Claude takes three steps, and all of the sign-in happens in the browser. These steps follow the SiteGPT MCP server docs.

Step 6: Add SiteGPT in Claude's connector settings

In Claude, open Settings, then Connectors. Find SiteGPT in the connectors directory and select Connect. The background is in SiteGPT in Claude's connector directory.

You can also add a custom connector with the URL https://sitegpt.ai/mcp.

Step 7: Approve access on the SiteGPT approval page

SiteGPT opens an approval page in the browser. It shows the app that asks for access, the permissions, and whether the app can reach all chatbots or only selected ones.

Under Access, you can select Read only. A read-only connection can list chatbots and read conversations, leads, analytics, content sources, and settings. It cannot create, change, or delete anything.

Select Connect. The connection can never do more than your dashboard role allows. Permissions that your role does not have are left out of the page.

Step 8: Check that it works

Ask Claude: "List my SiteGPT chatbots." Claude calls a SiteGPT tool and shows your chatbots.

To end access later, open Agents in the SiteGPT top navigation. The MCP connections section lists each connection with its permissions. Select Revoke to end one.

A worked example: one week of chatbot management from Claude

A worked example shows what the connection is for. The store below is hypothetical. It sells bike parts and runs one chatbot named Support Bot on the Growth plan.

Day	Prompt to Claude	What happens	Read or write
Monday	"List my SiteGPT chatbots, then show this week's unresolved conversations for Support Bot."	Claude searches the API catalog, then reads chatbots and open conversations	Read
Tuesday	"Which questions did the support bot fail to answer this week?"	Claude returns the unanswered questions so the owner can see what the content lacks	Read
Wednesday	"Retrain the bot on the new release notes and add a custom response for the pricing change."	Claude queues the sources and drafts the custom response. It asks for approval before the write	Write
Thursday	"List the new leads from Support Bot this week."	Claude reads the leads the chatbot collected	Read
Friday	"Check our SiteGPT usage and whether we are close to any plan limits."	Claude reads usage and limits from the account	Read

Four of the five prompts only read data. A read-only connection covers them. Only Wednesday's prompt needs write permission.

For more prompts of the same kind, see 10 customer service analytics questions you can ask Claude.

How to let an agent build a chatbot before you have an account

An AI agent can build a temporary chatbot from a website URL with no SiteGPT account. SiteGPT calls this agent-first onboarding. It needs an agent that can run terminal commands and Node 18 or later.

The agent installs the CLI with npm install -g @sitegpt/cli.

It runs sitegpt onboarding start with the website URL and the --json flag. This command needs no login.

The response includes a temporary token that works only for that chatbot. The agent uses it to add content, a persona, instructions, and appearance.

The agent checks progress with sitegpt onboarding status and fixes items marked pending.

The agent gives you the onboarding URL. You preview the chatbot there and claim it into an account. To claim it, you choose a plan at checkout.

The full flow is in the agent-first onboarding docs. The SiteGPT agents page shows how the CLI, MCP, and API fit together.

Summary graphic of four ways an assistant reaches a SiteGPT chatbot: remote MCP server, local MCP server, the CLI, and agent-first onboarding

How to choose a setup path

Choose a setup path by where your assistant runs. Claude on web and desktop is the confirmed MCP client for SiteGPT. SiteGPT also publishes setup guides for Claude Code, Cursor, and Codex.

If you need...	Pick...	Why
To manage a chatbot from Claude on web or desktop	The remote MCP server through Claude's connector settings	Nothing to install, and access is approved in the browser
To look at data without any risk of changes	The remote MCP server with Read only selected	Write calls fail on a read-only connection
To work from the terminal in Claude Code	The Claude Code setup guide	One command adds the remote server: `claude mcp add --transport http sitegpt https://sitegpt.ai/mcp`
To work inside the Cursor editor	The Cursor setup guide	The server URL goes in `.cursor/mcp.json`
To run it from Codex or a CI pipeline	The Codex setup guide	Codex starts the local server with `npx -y @sitegpt/cli mcp`
An agent limited to one chatbot and a few permissions	The CLI with a separate profile	`sitegpt login` accepts `--scope` and `--chatbot` options
To try a chatbot before signing up	Agent-first onboarding	No account or login is needed to start

A CLI login for a content-only agent looks like this:

sitegpt login --profile knowledge-agent --chatbot YOUR_CHATBOT_ID --scope knowledge:read --scope knowledge:write

The terminal route is covered step by step in the chatbot CLI guide.

Decision tree for choosing between a Claude project or custom GPT, an API build, and a content-trained platform, and between the MCP server and the CLI

Common mistakes when you build and connect the chatbot

Most problems with this setup come from six mistakes. Each one has a short fix.

Mistake	What you see	Fix
Connecting on the Starter plan	The approval page says API access needs the Growth plan or above. Calls fail with 403 PLAN_UPGRADE_REQUIRED	Move to Growth or above
Granting every permission when you only want reports	Claude can create, change, and delete	Select Read only on the approval page
Reconnecting to switch to read-only	The old connection keeps its permissions	Revoke the old connection on the Agents page, then connect again
Approving in the wrong browser profile	The approval page does not return to the app	Sign in to the right SiteGPT account in the same browser, and allow pop-ups
Pasting an API token into the chat	A credential sits in a conversation	Use the browser approval for the remote server, or a saved CLI profile
Skipping the test in step 4	Visitors get "not enough information" replies	Ask real questions first and add the missing pages

One limit applies to healthcare teams. For a chatbot in a HIPAA workspace, the MCP servers do not return conversations, leads, or escalations. Those stay in the SiteGPT dashboard. A HIPAA BAA is available on the Enterprise plan only, and the HIPAA docs describe how a covered workspace behaves.

SiteGPT has other limits to plan for. There is no free plan. There is no Intercom or HubSpot integration today. Message quotas are monthly, and one GPT-4.1 message uses as much quota as ten GPT-4.1 mini messages.

FAQ

Building the chatbot

Can you build a custom AI chatbot with ChatGPT?

Yes, in two ways. A Business, Enterprise, or Edu workspace can create a custom GPT that runs inside ChatGPT. A website chatbot needs the OpenAI API or a platform that uses OpenAI models, such as SiteGPT.

Can you build a chatbot with Claude?

Yes. A Claude project gives a team a workspace with its own knowledge base and instructions inside Claude. For a website chatbot, a developer can build on the Anthropic API, or Claude can manage a chatbot platform over MCP.

Do you need to code to build the chatbot in this guide?

No. Steps 1 to 5 use the SiteGPT dashboard, and steps 6 to 8 use Claude's settings and a browser approval page. Only the agent-first onboarding path uses terminal commands. Other no-code options are compared in the best no-code chatbot builders.

MCP and Claude

What is an MCP chatbot?

An MCP chatbot is a chatbot on a platform that exposes an MCP server. An assistant such as Claude connects to that server and manages the chatbot for the owner. A longer walkthrough is in how to build and run a chatbot from Claude.

What is the SiteGPT MCP server URL?

The URL is https://sitegpt.ai/mcp. Add it as a custom connector in Claude, or find SiteGPT in Claude's connectors directory.

Do you paste an API token into Claude?

No. The approval page creates an OAuth grant for the app. You can revoke the grant at any time on the Agents page.

Can Claude do more in SiteGPT than you can?

No. The connection is limited to what your dashboard role allows. You can narrow it further to read-only access or to selected chatbots.

Cost and limits

How much does SiteGPT cost with MCP access?

MCP access starts on Growth: $129 a month billed monthly, or $79 a month billed yearly. Scale costs $429 a month billed monthly, or $259 a month billed yearly. Enterprise pricing is custom.

Is there a free way to try it?

There is no free plan. Every self-serve plan has a 7-day free trial that requires a credit card. An agent can also build a preview chatbot with no account through agent-first onboarding.

Is an FAQ-only chatbot a simpler place to start?

Yes, if the questions are few and fixed. How to build an FAQ chatbot compares four ways to do that.

Build the chatbot, then connect Claude

Start with steps 1 to 5 during the trial, and connect Claude once the chatbot answers your test questions well. Start a SiteGPT free trial on the Growth plan if you want MCP access from day one.

Sources

Checked on October 7, 2026.

Model Context Protocol documentation: What is MCP?

Anthropic: Introducing the Model Context Protocol

Claude Help Center: Get started with custom connectors using remote MCP

Claude Help Center: What are projects?

OpenAI Help Center: Creating and editing GPTs

SiteGPT docs: MCP server

SiteGPT docs: MCP tool reference

SiteGPT docs: Plans and limits

SiteGPT docs: CLI authentication

Building Custom AI Chatbots With ChatGPT and Claude: How to Build One and Manage It From Claude