SiteGPT is listed in Claude's connector directory, at claude.ai/directory/connectors/sitegpt. That is a stronger claim than "SiteGPT has an MCP server", because anyone can check it: open the directory in Claude, search for SiteGPT, and the Connect button is there. No URL to paste, no custom connector form.
This post is the short version of what that means and what to do with it. It is deliberately not a setup walkthrough; the full guide to building and running a chatbot from Claude already exists.
iShort answer
Claude users can now add SiteGPT from the official connector directory in two clicks: find the listing, approve access over OAuth in the browser. The connector is SiteGPT's remote MCP server with its three-tool interface: search finds the right operation, execute_read reads data, execute_write changes things and asks before doing so. Permissions are a subset of the connecting user's own dashboard role, never more. A listing is a distribution fact, not an endorsement; the checkable parts are the OAuth flow and the permission model.
Two clicks of setup, then the connector is a control surface for your live chatbot.
Find the listing
In Claude, open the connector directory and search for SiteGPT, or go to claude.ai/directory/connectors/sitegpt and click Connect.
Approve over OAuth
A browser window signs you in to SiteGPT and asks what to grant: all chatbots or only selected ones. No API keys touch the chat.
Ask the first question
Try: "List my chatbots and their status." "Pull yesterday's leads and draft follow-ups." "Which questions did the bot fail to answer this week?"
What being in Claude's connector directory means
Claude's directory is the catalog of connectors users can browse and add from inside Claude. Connectors are integrations built on MCP, the open Model Context Protocol, which is how Claude talks to outside tools. Until now, connecting SiteGPT meant adding https://sitegpt.ai/mcp as a custom connector by hand; a directory listing removes that step and puts SiteGPT where Claude users already look for tools.
The listing itself is short: SiteGPT, described as "Manage AI support chatbots from your AI assistant", carrying the directory's Community label. That label is worth reading correctly, and it leads to two honest framings, because this is easy to oversell. A listing is distribution, not endorsement: it means the connector is available in the catalog, not that Anthropic certifies SiteGPT. And the connector itself is unchanged; the directory is a shorter path to the same remote MCP server, with the same OAuth approval and the same permission limits.
Adding SiteGPT to Claude, the two-click version
First, find the listing: open the connector directory in Claude and search for SiteGPT, then click Connect. Second, approve: a browser window opens, you sign in to your SiteGPT account, and the approval screen asks whether the grant covers all your chatbots or only selected ones. That is the whole setup.
The OAuth detail worth knowing is what does not happen: no API key is generated, pasted into a chat, or stored by the client. The server advertises its OAuth metadata automatically and clients register through Dynamic Client Registration, which is why there is nothing to configure. Disconnecting the connector in Claude's settings ends its access.
What the connector can do on day one
The interface is three tools, and knowing them explains every capability and every limit.
| Tool | What it does | What that means in practice |
|---|---|---|
| search | Finds the right operation in the SiteGPT API v2 catalog | Claude discovers what is possible instead of guessing |
| execute_read | Runs read-only calls through an authenticated bridge | List bots, pull conversations and leads, check usage |
| execute_write | Creates, updates, and deletes, marked destructive | Update knowledge, edit personas, always with a confirmation step |
In practice, day one looks like asking. "List my chatbots and their status." "Pull yesterday's leads, qualify them, and draft follow-up emails." "Add the new refund policy page to the bot's knowledge." "Which questions did the bot fail to answer this week, and what should the missing help article say?" The first three are reads and writes against your live workspace; the last one pairs a read with the drafting work an assistant is good at. More of what this changes for support teams is at MCP for customer service.
What it cannot do, and why the permissions work that way
The grant is a subset of you, never a superset. The client cannot receive permissions or chatbot access your dashboard user does not have, so a read-only teammate's Claude cannot edit knowledge, and an approval scoped to one chatbot cannot touch the others.
Anything destructive goes through execute_write, which is marked destructive precisely so clients confirm before acting. And the connector operates your workspace rather than replacing your judgment: it will not answer your customers for you (the widget on your site keeps doing that), and it cannot see anything your own login could not.
Where to go next
If the directory listing is how you found SiteGPT, the product behind it is a customer service chatbot trained on your own content, from $39/mo on the annual Starter plan at SiteGPT's pricing, with a 7-day free trial.
If you are already a customer, the first thing to do with the connector is ask it questions: ten support analytics questions to ask Claude about your chatbot has the exact prompts, from "which questions did the bot fail to answer this week" to a Monday summary written for your Slack channel.
Beyond that, the connector is one member of a family. The agents page is the overview of what an assistant can operate end to end. The same MCP server works from your tools: Claude Code in the terminal, Cursor in the editor, and Codex in CI each have a setup guide, and the SiteGPT CLI covers scripted runs and a local MCP server. For standing agents that hold a role rather than answer one prompt, see OpenClaw and Hermes.
Frequently asked questions
What is Claude's connector directory? A catalog inside Claude where users browse and add connectors, the integrations that let Claude work with outside tools over MCP, the open Model Context Protocol. Instead of pasting a server URL as a custom connector, you find the tool in the directory and click Connect. SiteGPT's listing is at claude.ai/directory/connectors/sitegpt, and the underlying connector is the same remote MCP server documented at sitegpt.ai/mcp-server.
Does being listed mean Anthropic endorses or certifies SiteGPT? No. A directory listing is a distribution fact: it means the connector is available inside Claude's catalog and users can add it without manual setup. It is not an endorsement, a certification, or a partnership claim, and this post does not make one. The security-relevant facts are the ones you can check: OAuth in the browser, no API keys in chat, and permissions scoped to the connecting user's own dashboard role.
What can Claude do with SiteGPT once connected? The connector exposes three tools. search finds the right operation in the SiteGPT API v2 catalog. execute_read runs read-only calls, so Claude can list chatbots, pull conversations and leads, and check usage. execute_write handles creates, updates, and deletes, and it is marked destructive so clients ask for confirmation before changing anything. In practice that covers listing bots, reviewing leads, updating knowledge, and editing personas from the conversation.
What permissions does Claude get over my SiteGPT account? Only yours, at most. The connection is approved over OAuth in the browser, and the grant is scoped to the connecting user: the client cannot receive permissions or chatbot access your dashboard user does not have. The approval can also be scoped to all chatbots or only selected ones, and disconnecting the connector ends its access. A read-only teammate's Claude stays read-only.
Do I need Claude to use SiteGPT? No. The dashboard, the Agent API v2, and the CLI all operate the same chatbot, and the widget keeps answering visitors on your website either way. The connector adds a control surface for the work you would otherwise do by clicking through screens: checking leads, updating knowledge, reviewing what the bot could not answer.
Does this work with assistants other than Claude? The server is built on the open MCP standard, and Claude is the confirmed client today. SiteGPT publishes setup guides for driving it from Claude Code, Cursor, Codex, OpenClaw, and Hermes, and lists an assistant as confirmed only after the connection test passes. The ChatGPT plugin is not available.
Sources
- SiteGPT MCP server page for the directory listing statement, the three tools and their descriptions, the server URL, Dynamic Client Registration, and the subset permission wording, read 14 August 2026
- Claude connector directory listing for the listing name, description, and Community label, verified from a signed-in Claude session on 14 August 2026
- Model Context Protocol for the open standard the connector is built on
- SiteGPT pricing for plan rates and the 7-day trial
Last updated: August 2026.