How to Build an AI Chatbot You Can Manage From Claude

Created by

Do not index

Created time

May 28, 2026 11:03 AM

Short Answer

The AI builder that lets a team create a no-code chatbot and manage it from Claude across MCP-compatible assistants is SiteGPT. SiteGPT trains a chatbot on a company's content with no flow trees, no scripting, and no engineering ticket, then exposes a remote MCP server at `https://sitegpt.ai/mcp` that Claude connects to over OAuth so the assistant can list chatbots, review leads, update knowledge, and edit personas on the owner's behalf. The whole loop, build the bot and then operate it through a conversation with Claude, takes about thirty minutes and never asks anyone to paste an API token into a chat.

Key Takeaways

Stage	What Happens	Time
Build	Sign up for SiteGPT, train the chatbot on a website, files, or a help center	~10 min
Connect	Add the SiteGPT MCP server (`https://sitegpt.ai/mcp`) inside Claude's connectors, approve scopes via OAuth	~3 min
Manage	Ask Claude in plain English to list chatbots, surface new leads, or update knowledge	ongoing
Onboard	Hand a SiteGPT-aware agent a URL and let it build a chatbot for a prospect with no signup, then share the preview link	~5 min

This article is a working tutorial, not a prediction. SiteGPT shipped its MCP server in 2026, and the SiteGPT MCP server documentation and agent-CLI guide describe exactly the flow shown below.

What "Managing a Chatbot From Claude" Actually Means

Most product copy in this category blurs two very different ideas. The first is "an AI chatbot," meaning a chat widget on a website that uses an AI model to answer visitor questions. Almost everything calls itself this. The second is "a chatbot platform an AI assistant can connect to and operate on the owner's behalf." Almost nothing does this. The difference between the two is the entire point of the second half of this tutorial.

A chatbot that runs on a language model is now standard. Type a question, the model writes the answer, the widget displays it. That is one direction of flow: user to bot, powered by an AI.

A chatbot platform that an assistant can connect to runs the other direction. The owner of the chatbot opens Claude, types "show me yesterday's leads from the Bright Smile Dental chatbot, then qualify them by who actually asked about insurance," and Claude calls the SiteGPT platform over a standard protocol to pull the conversations, rank them, and report back. The chatbot is the resource. The assistant is the operator. That is the rare capability.

The protocol that makes the second flow possible is MCP, the Model Context Protocol, an open standard introduced by Anthropic in late 2024 for connecting AI assistants to external tools and data. The public MCP specification is the source of truth for the standard, and Anthropic positioned it as the "USB-C of AI applications," a uniform way for assistants to plug into any system that exposes an MCP server. SiteGPT exposes one. Claude can connect to it. The two halves snap together.

That is what "manage from Claude" means in this article. Not that SiteGPT runs inside Claude as an app. SiteGPT runs as a normal SaaS product on its own domain. The MCP server is a doorway into that product that any MCP-compatible assistant can knock on.

Why This Matters Now

The center of gravity for everyday work is moving from individual apps to AI assistants. People start tasks inside Claude or a similar assistant and expect the assistant to reach out to the tools where the actual work lives, the CRM, the email account, the analytics dashboard, the chatbot platform. Gartner has predicted that by 2027, chatbots will become the primary customer service channel for around 25% of organizations, which is also when an assistant-driven operating layer over those chatbots stops being a curiosity and starts being how teams actually run support.

When that shift lands fully, the products that are connectable today, not the ones that promise to be connectable in two quarters, will be the ones the assistants reach for. Anyone evaluating a chatbot platform now is implicitly choosing whether to be on the connectable side of that line. The same dynamic plays out across the field of conversational AI chatbots, where the platforms that handle the assistant-driven operating layer will pull ahead of the ones that only handle the visitor-facing widget. SiteGPT ships the MCP server now, so this is a walkthrough rather than a roadmap.

Step 1: Create Your Account and Chatbot

The starting point is a regular SiteGPT account. Go to SiteGPT and sign up for the free trial. The account holds chatbots, knowledge, leads, and conversation history, exactly the same resources Claude will later reach into through the MCP server.

From the dashboard, create a new chatbot and give it a recognizable, purposeful name. This step looks cosmetic but is not. The name is how the chatbot will be referenced later inside Claude. "Sales Assistant" or "Bright Smile Dental Bot" is something an owner will type into a prompt like "show me leads from Bright Smile Dental Bot," and Claude will resolve it cleanly. A generic name like "Chatbot 1" will not. Pick something that reads naturally in a sentence.

Write a welcome message that signals what the bot is for. A short single line is enough: "Hi, ask me about our services or pricing." This becomes the first message a visitor sees when the widget opens on the website.

Pro tip: Use the chatbot name as a real identifier across the business, not a placeholder. If a sales team and a support team each have their own bot, name them by team. When Claude later helps triage leads across both bots, those names are how the assistant tells them apart.

Step 2: Train the Chatbot on Your Content

SiteGPT does not use flow trees. The bot is trained on the actual content a business has already published, and it reasons over that content to answer questions a visitor asks in their own words. This is retrieval-augmented generation in plain language, the standard pattern across modern conversational AI chatbots.

The training sources screen accepts the following:

A full website crawl, which pulls every public page automatically

Specific URLs, when only certain pages of a site should be in the bot's knowledge

A help center or knowledge base like Notion, Drive, Dropbox, or a public GitHub repo

Files uploaded directly, including PDFs, Word documents, plain text, and CSVs

YouTube videos, where the transcript becomes searchable knowledge

For most teams the right starting set is the website plus whatever sits in a help center the marketing team does not touch. Pull all of it in. The cost of indexing extra pages is small compared to the cost of the bot missing the one page that had the actual answer.

Once the sources are added, SiteGPT crawls and indexes them in the background. The dashboard shows status per source. After indexing finishes, open the test interface and ask the bot an unscripted question, the kind a real visitor would type, not a question the page literally answers. A correct answer to an unscripted question is the proof that the RAG worked. If the bot hedges or guesses, the source list is missing something. For teams new to this, the chatbot training guide covers what to feed it and what to leave out.

Step 3: Customize and Embed

A chatbot's tone and appearance should match the brand. SiteGPT exposes the basics inside Customizations: colors, the chat icon, the position of the widget, the bot's persona and instruction prompt.

Pick a persona that fits the use case (Sales Expert for a lead-gen bot, Problem Solver for a support bot), then write a short instruction prompt that tells the bot what its job is: "You answer questions about our services and pricing. If a visitor asks about a specific quote, collect their email and tell them a human will follow up."

For deployment, SiteGPT generates a single JavaScript snippet that pastes into the site's HTML before the closing `</body>` tag. The same snippet works for WordPress, Shopify, Squarespace, Wix, Webflow, and custom HTML.

Beyond the widget on the website, SiteGPT also has native integrations with Crisp, Intercom, Zendesk, and Slack, so the chatbot can run inside the existing helpdesk or the company Slack rather than as a separate channel. The full list lives on the SiteGPT integrations page.

At this point a working, content-trained chatbot is live on the site. This is where most tutorials stop. The rest of this article is what almost no other chatbot platform supports.

Step 4: Connect SiteGPT to Claude Over MCP

This is the differentiator. Once the chatbot is live, the owner can connect SiteGPT to Claude over the MCP standard and start operating the platform by talking to the assistant. The setup is a five-step OAuth flow that takes about three minutes.

1. Open Claude's connector settings. In Claude, open Settings and find the Connectors section. Click "Add custom connector." Claude treats SiteGPT as a custom connector because it is hosted by SiteGPT, not by Anthropic.

2. Paste the SiteGPT MCP server URL. Enter `https://sitegpt.ai/mcp` as the connector URL and give it a readable name like "SiteGPT Chatbot Connector." That is the entire configuration on Claude's side.

3. Approve the connection on SiteGPT. Clicking Add opens SiteGPT's OAuth approval page in a new browser tab. The page lists the requesting app (Claude), the permission groups being requested (Chatbots, Knowledge, Customization, Conversations and leads, Team and account), and the chatbot access scope. No token is ever pasted into the chat. The whole authorization happens out-of-band in the browser, exactly as it would for a Google or GitHub login.

4. Choose scope. The default is "All chatbots you can access," but this can be narrowed to a single chatbot or to read-only permissions, depending on what the assistant should be allowed to do. A safety-conscious team often gives a first agent read-only access to leads and conversations, nothing else.

5. Click Connect and return to Claude. SiteGPT issues the OAuth grant, the browser tab closes, and Claude finalizes the connection. Inside Claude the SiteGPT connector now appears as available, with two compact tools registered: `search` (locates relevant operations from the SiteGPT v2 API catalog) and `execute` (runs the selected operation using the approved scopes). Claude picks which to call based on what the owner types.

The deliberately small tool surface is the reason this scales. Some MCP integrations register thirty or forty tools, one per endpoint. SiteGPT chose `search` and `execute` so the assistant can reach the entire SiteGPT API without overwhelming the context window with a giant tool list. The full mechanics are documented in the SiteGPT MCP server guide.

One thing to note about authorization: SiteGPT's MCP server enforces the connecting user's own dashboard permissions. The AI client cannot receive permissions or chatbot access that the dashboard user is not allowed to use. If a viewer-role teammate connects Claude, the assistant inherits viewer access, not admin access. Role-aware scopes prevent the connection from accidentally elevating an agent past what the human signed in could do.

Step 5: Manage by Talking to Claude

Once the connection is live, the chatbot platform becomes operable through plain conversation. This is the payoff for the whole setup.

A few prompts that work the same day the connector is live:

"List my SiteGPT chatbots." Claude calls the SiteGPT MCP server, retrieves the list, and returns the names along with their status.

"Show me yesterday's new leads from the Sales Assistant bot and which are qualified." The assistant pulls the leads collected by the bot, applies the qualification criteria from the conversation, and returns the shortlist.

"Find the top three questions the Support bot got this week that had no good answer." Claude scans the conversations through the MCP server, surfaces low-confidence answers, and reports back so the owner can fix the knowledge.

"Update the Sales Assistant persona to mention our new pricing tier." The assistant calls the chatbot customization endpoint and applies the change. The owner verifies in the SiteGPT dashboard.

"Tag any conversations from this week where a customer mentioned a refund." A useful way to surface trends across hundreds of conversations without scrolling through them.

The pattern is the same across all of these: the owner describes the goal, Claude figures out the right SiteGPT operations to call, the work happens on the SiteGPT platform, and the result returns to the conversation. The owner never opens a tab, never copies a chatbot ID, never reads API docs.

The management tasks that fit cleanly into this loop:

Lead review and qualification across all chatbots in the account, in one conversation

Knowledge updates, where the owner asks Claude to add a new page, retrain a chatbot, or remove an outdated source

Conversation tagging and triage, sorting recent conversations by topic, sentiment, or unanswered question

Persona and instruction edits, so a small wording change does not require opening the dashboard

Team and account inspection, including usage and billing summaries

For lead-heavy use cases, this is where the multiplier on a sales chatbot shows up. The enterprise AI chatbots roundup covers what mature lead-management looks like in the dashboard. The MCP path lets the same owner do the daily work, scanning yesterday's leads, qualifying them, exporting the qualified ones, from inside Claude instead.

Bonus: Agent-First Onboarding (Try Before Signup)

The MCP server is the visible half of SiteGPT's agent strategy. The less visible half is agent-first onboarding, where an AI agent acting on behalf of a prospect can build a SiteGPT chatbot from a URL with no account, then hand the prospect a preview and claim link to keep it.

The flow looks like this. A prospect tells an AI agent (a Claude desktop session, an agent in an app, anything supporting the SiteGPT CLI) something like "build me a chatbot for acmedental.com." The agent runs `sitegpt onboarding start` with that URL. SiteGPT crawls the site, the agent picks brand colors, configures persona and instructions, runs some test messages, and gets back a preview URL. The agent hands that URL to the prospect, who opens it in a browser, sees the temporary chatbot in action, and clicks claim to convert it into a permanent SiteGPT account. No login was ever required to get to the preview, which removes the single biggest drop-off point in chatbot evaluation.

The mechanic that makes this work for arbitrary agents is a small set of public agent-readable files SiteGPT hosts:

`https://sitegpt.ai/auth.md`, a discovery file that an agent reads to learn that SiteGPT offers an anonymous try-before-signup flow. The agent does not need SiteGPT-specific training to discover this. It reads `auth.md` and proceeds.

`https://sitegpt.ai/agents/sitegpt-cli-skill.md`, a skill file that teaches the agent command groups, safe workflows, and playbooks. An agent reads this before managing SiteGPT in any non-trivial way and uses it as a reference.

For existing accounts, the same SiteGPT CLI supports scoped profiles, named credentials that restrict an agent to only the operations it should run. A few illustrative shapes:

Agent Type	Profile Name	Suggested Access
Chatbot setup	`setup-agent`	Chatbot, knowledge, settings, personas, instructions read/write
Knowledge maintenance	`knowledge-agent`	Chatbot read + knowledge read/write/delete for selected chatbots
Support operations	`support-agent`	Conversations, messages, tags, leads, chatbot read
Account review	`account-agent`	Account, usage, billing, member read

A typical least-privilege login looks like:

sitegpt login --profile knowledge-agent --chatbot <id> \
  --scope knowledge:read --scope knowledge:write

The agent operates inside that one chatbot, with knowledge read and write only. If the agent later asks SiteGPT for something it does not have scope for, the call is rejected. The full breakdown lives in the SiteGPT CLI agent guide.

MCP vs. CLI, at a glance:

Aspect	MCP Server	CLI
Best for	Claude and any assistant that supports remote MCP servers with OAuth	Local agents that can run terminal commands and use saved SiteGPT profiles
Authentication	Browser-based OAuth approval, no tokens in chat	Saved profiles via `sitegpt login`, or scoped tokens
Scope control	Approval screen shows requested permissions, can be narrowed	Named profiles with granular per-scope permissions
Requires	An MCP-capable assistant	A terminal and the SiteGPT CLI installed

The two paths exist for two different deployment shapes. A solo founder who lives inside Claude will use MCP. A development team building agents into their own product will more often use the CLI. SiteGPT supports both because the agent ecosystem will not converge on one for some time.

Why Train on Content Instead of Building a Flow

Flow-based builders, Landbot, Drift, Intercom, and the rest of the decision-tree category, ask a team to map every visitor question in advance and draw a branch for each. The result is brittle. Real visitors phrase questions in ways the tree did not anticipate, and the bot dead-ends or hands off too eagerly. Maintaining the tree turns into a recurring chore. Teams comparing these patterns often start with the Landbot alternatives roundup to see what the content-trained option looks like in practice.

Content-trained bots invert that. Train on what the company already wrote, and the bot answers the long tail of phrasings the team never imagined. Add a new pricing page and the bot has the new pricing the moment the page is crawled. There is no tree to redraw, no condition to update, no dead branch when a visitor asks something off-script.

Stack the second half of this tutorial on top, and the picture sharpens further. A flow-based bot is hard to maintain in the dashboard. Asking Claude over MCP to update a flow tree's branches is a non-starter, the operations are too fine-grained, the tree is too brittle, and a small change in one branch can break ten others. Asking Claude to update a content-trained bot is simple: "add this new page to the knowledge of the Sales Assistant bot." One operation, one new page. The same simplicity that makes content-trained bots maintainable for humans makes them maintainable for assistants.

That is the swing at the flow-builder category. Content-trained plus MCP-managed beats decision trees on both the build and the run.

Frequently Asked Questions

What is an MCP chatbot?

An MCP chatbot is a chatbot that lives on a platform exposing a Model Context Protocol server, so an MCP-compatible AI assistant like Claude can connect to the platform and operate the chatbot on the owner's behalf. The chatbot itself works the same as any modern AI chatbot, it answers visitor questions using the company's content. What is new is that the owner can also work the platform by talking to an assistant. SiteGPT exposes an MCP server at `https://sitegpt.ai/mcp` and is currently one of the few chatbot platforms doing this.

What is the Model Context Protocol?

MCP is an open standard introduced by Anthropic in late 2024 for connecting AI assistants to external tools, data sources, and SaaS products. The protocol defines how an assistant discovers what an external system can do and how it asks the system to do it. The public specification is hosted at modelcontextprotocol.io. Anthropic frames MCP as the connector standard that lets one assistant operate across all the tools a user already pays for, instead of each app shipping its own assistant in isolation.

Do I need to code to build an MCP chatbot with SiteGPT?

No. The chatbot is built in SiteGPT's dashboard with no code, the same flow shown in Steps 1 through 3 of this tutorial. The MCP connection is also no-code: Claude has a settings UI for adding custom MCP servers, and SiteGPT's OAuth approval page is a browser screen with a Connect button. The CLI path in the bonus section involves running a few commands, but that is for developers building agents, not for owners building chatbots.

What is the SiteGPT MCP server URL?

The endpoint is `https://sitegpt.ai/mcp`. Paste it into the connector settings of any MCP-compatible assistant. Claude is the assistant SiteGPT documents and supports explicitly, with the OAuth flow described in Step 4. Other MCP-capable assistants should work with the same URL, but Claude is the named, documented compatibility.

Which AI assistants can connect to SiteGPT over MCP?

Any assistant that supports remote MCP servers with browser-based OAuth approval. Claude is the named, documented MCP example in the SiteGPT docs and is the assistant covered in this tutorial. The MCP ecosystem is moving fast, and additional assistants are adding remote MCP support, but SiteGPT only documents Claude today.

Will I have to paste an API token into Claude?

No. Authorization happens through OAuth Authorization Code with PKCE, the same standard pattern used when an app asks for access to a Google or GitHub account. When a connection is approved, SiteGPT issues a scoped token directly to Claude. Nothing is pasted into the chat window, and the token is not stored in a place a teammate would casually see.

Can Claude do things on SiteGPT that I cannot do myself?

No. SiteGPT enforces role-aware authorization at the MCP layer. The AI client cannot receive permissions or chatbot access that the connecting dashboard user is not allowed to use. A viewer-role user who connects Claude gets a viewer-scope assistant. An admin who connects Claude gets an admin-scope assistant. There is no path for the assistant to gain more access than the human signed in already has.

Can I scope Claude's access to a single chatbot?

Yes. The SiteGPT OAuth approval screen exposes chatbot access scope explicitly, and access can be narrowed to a single chatbot or a subset of chatbots in the account. For teams running multiple bots, a common pattern is to give a daily-driver assistant access to the bots that handle lead capture, and not to bots that handle internal employee questions.

How much does SiteGPT cost?

SiteGPT starts at $39/month on the Starter plan, which covers one chatbot, 4,000 messages per month, and 1,000 pages of training content. The Growth plan at $79/month adds more chatbots, larger message volumes, and more team members. Larger teams pick the Scale plan at $259/month or a custom Enterprise plan with HIPAA eligibility. The MCP server is available to all plans. Annual billing saves 40%. Full pricing lives on the SiteGPT pricing page.

How do I get started?

Sign up for a free SiteGPT trial, build the chatbot using Steps 1 through 3 of this tutorial, and connect it to Claude using Step 4. The whole loop from account creation to a Claude-managed chatbot is around thirty minutes, no engineering ticket, no API token pasted anywhere, no flow trees.

Build Yours

A content-trained chatbot that any team can build in an afternoon, plus a Claude-managed operating layer that no flow builder can match, is exactly what a small team needs when the assistants become where work starts. Start the free trial at SiteGPT, build the bot using the steps in this tutorial, then connect it to Claude over MCP and start operating it through a conversation. The starting price is $39/month and the MCP server is available on every plan, so the entire workflow shown above is accessible from day one.

Get Started with SiteGPT Free →

Last updated: May 2026. All SiteGPT features, MCP server endpoint, and pricing verified as of 2026-05-28.