Skip to main content
SiteGPT provides an MCP server for supported AI assistants. The assistant connects to SiteGPT, asks you to approve access in the browser, and then uses SiteGPT tools on your behalf. Use the MCP server when your AI assistant supports remote MCP servers and OAuth. Use the CLI when the assistant runs on your machine and can execute terminal commands.

​
Endpoint

Use this MCP server URL:
When a client connects without credentials, SiteGPT advertises OAuth metadata so the client can open the approval flow automatically. Discovery itself needs no credentials: an anonymous client can complete the initialize handshake, list every tool, and read the app-view resources — useful for connectivity checks and directory scanners. Every tools/call requires the OAuth grant. The server’s machine-readable card is published at https://sitegpt.ai/.well-known/mcp.json.

​
Connect Claude

1

Open Claude's integrations or connectors settings

Open the place where Claude lets you add a custom MCP server.
2

Add the SiteGPT MCP server URL

Enter:
3

Start the connection

Claude should open SiteGPT’s OAuth approval page in your browser.
4

Approve SiteGPT access

Review the app name, permissions, and chatbot access. Click Connect only if the request looks correct.
5

Return to Claude

After approval, Claude completes the connection and shows the SiteGPT MCP tools.

​
What Claude can do

The MCP server exposes 17 tools in two layers: three API tools that can reach the entire SiteGPT API v2, and purpose-built tools for common jobs that pair with inline app views.

​
API tools

Read and write are separate tools with explicit annotations, so AI apps can group them correctly in their permission UI — reads can be allowed always while writes stay behind per-call approval. SiteGPT still enforces your dashboard role, token scopes, and chatbot restrictions on every call.

​
Purpose-built tools

Common jobs have first-class tools that return compact results and carry the same explicit read/write annotations:

​
Inline app views

On AI apps that support the MCP Apps extension, the purpose-built tools render as live views directly in the conversation instead of raw JSON. Eight views ship: the actual chat widget (its built-in chat runs through send_chat_message), an onboarding progress card, an appearance card, a conversations inbox with a transcript panel, an analytics card, a leads browser, an escalations queue, and a knowledge sources view. The same views are also published in the OpenAI Apps format, so hosts that use that dialect render them too. On clients without app-view support, every tool still works and returns its compact structured result.

​
Approval and permissions

The approval page creates an OAuth grant for the AI client. It does not ask you to paste an API token into Claude. During approval, SiteGPT shows:
SiteGPT’s authorization is role-aware. The AI client cannot receive permissions or chatbot access that your dashboard user is not allowed to use.

​
MCP Server vs CLI

You can use both. For example, connect Claude through the MCP server for interactive work, and keep a local sitegpt CLI profile for scripts or local agents.

​
Test the connection

After connecting, ask the AI assistant:
Then try a more specific request:
The assistant should use the SiteGPT MCP tools to search for the right API operation and call it with execute_read or execute_write.

​
Troubleshooting

​
The approval page opens but does not return to the AI app

Wait a few seconds first. Some MCP clients take a moment to finish the OAuth callback. If it stays stuck:
  • Make sure third-party cookies, popup blockers, or browser privacy extensions are not blocking the OAuth callback.
  • Cancel the connection in the AI app and try adding https://sitegpt.ai/mcp again.
  • Confirm you are signed in to the correct SiteGPT account in the same browser.

​
The AI app shows fewer tools than expected

The full surface is 17 tools: search, execute_read, and execute_write for the whole API, plus the purpose-built tools listed above. The tool list is available even before authorization, so if the AI app shows no SiteGPT tools at all, the connection itself failed — remove and re-add the server.

​
The AI app says a SiteGPT API call failed

Ask the assistant to show the exact endpoint, status code, and error message. Common causes are missing permissions, chatbot restrictions, or an expired/revoked grant. You can also verify your SiteGPT account from the CLI:

​
The CLI works but MCP does not

The CLI and MCP server use different authentication flows:
  • CLI: OAuth device login or saved SiteGPT API token.
  • MCP Server: OAuth Authorization Code + PKCE through the MCP client.
If CLI works, your SiteGPT account and API are healthy. Reconnect the MCP server in the AI app so it can create a fresh OAuth grant.

​
The docs MCP server

This documentation has its own MCP server, separate from the account server:
It is hosted by Mintlify, requires no authentication, and exposes search and retrieval tools over every docs page, so an assistant can look up how SiteGPT works while it works on your account. Its machine-readable descriptor is published at https://sitegpt.ai/.well-known/mcp/docs-server-card.json. Connect both servers: the account server to act on your chatbots, the docs server to answer questions about SiteGPT behavior without leaving the conversation.

​
Where to find the server

Both SiteGPT servers are listed in the official MCP registry:

​
Run the server locally (stdio)

Clients that prefer a local stdio server over the remote endpoint can run the npm launcher:
It delegates to sitegpt mcp from @sitegpt/cli — one implementation, two entry points. Authenticate with a SITEGPT_API_TOKEN environment variable or the CLI’s login flow. The hosted endpoint stays the recommended path: nothing to install, and OAuth approval happens in the browser.