Elfsight AI Chatbot — MCP Server

This repository holds the documentation and the registry manifest (server.json) of the Elfsight AI Chatbot MCP server. The server itself is hosted by Elfsight; there is nothing to build or run here. For help, contact Elfsight Help Center.

A remote Model Context Protocol server over your Elfsight AI Chatbot widget data. Point an MCP-capable AI agent at it and it can read your conversations, leads, knowledge base and analytics — and add answers to your knowledge base — without you copy-pasting anything into a chat window.

The server is account-scoped: once you sign in, the agent can reach every AI Chatbot widget on your Elfsight account, and nothing else (no other Elfsight products or data).

Server URL: https://widget-data.service.elfsight.com/mcp/ai_chat (Streamable HTTP).

Connect

Sign-in is OAuth. Add the server URL to your MCP client; the client opens a browser window, you sign in to Elfsight and approve access. There is no key to copy.

Claude Code

claude mcp add --transport http elfsight-ai-chatbot https://widget-data.service.elfsight.com/mcp/ai_chat

Then run /mcp inside Claude Code and choose Authenticate for elfsight-ai-chatbot.

Other clients

Any client that supports remote MCP servers with OAuth (dynamic client registration and PKCE) should connect with the server URL alone. We have tested Claude Code only; the snippets below are the usual shape of each client's config, not something we have verified.

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "elfsight-ai-chatbot": {
      "url": "https://widget-data.service.elfsight.com/mcp/ai_chat"
    }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "elfsight-ai-chatbot": {
      "type": "http",
      "url": "https://widget-data.service.elfsight.com/mcp/ai_chat"
    }
  }
}

Windsurf (mcp_config.json):

{
  "mcpServers": {
    "elfsight-ai-chatbot": {
      "serverUrl": "https://widget-data.service.elfsight.com/mcp/ai_chat"
    }
  }
}

What you can ask it

• list_assistants — "Which AI Chatbot widgets do I have?"
• get_analytics — "How did my AI Chatbot perform over the last 30 days?"
• search_conversations — "Show me conversations from last week about shipping."
• get_conversation — "What happened in conversation <id>?"
• list_leads — "List the new leads from this week that I haven't reviewed yet."
• search_knowledge — "Does my knowledge base say anything about refunds?"
• list_knowledge_gaps — "What do visitors ask that my assistant can't answer?"
• answer_knowledge_gap — "Add this as a knowledge base answer and close that gap."

answer_knowledge_gap is the only tool that writes. Three ready-made prompts wrap common workflows end to end: weekly_report, knowledge_gaps_review, lead_followup.

Privacy

Everything a tool returns can be sent to your AI provider (Anthropic, OpenAI, etc.) as part of your conversation with the agent. Keep that in mind:

• Lead contact fields are masked by default. A lead's name, email and phone come back masked (e.g. j***@example.com, +7***4567). Pass contacts: "full" on get_conversation or list_leads to get the real values in that call.
• Masking covers those fields only. Message text, page URLs and attachment links are returned as they are. Visitors often type their email or phone into the chat, so get_conversation can show contact details even when the lead fields are masked. Treat everything a tool returns as possibly containing personal data.
• Attachments are links, not content. File attachments come back as URLs, so a tool call never puts a file's contents into the model's context.

Internal and test conversations are not filtered out of any tool — if you use your own widget for testing, expect to see that traffic in the results too.

Limits and rate limiting

• search_conversations, list_leads, search_knowledge and list_knowledge_gaps return 20 items by default, 100 at most. A truncated result says so and reports the total. search_conversations and list_leads page with cursor; for the other two, narrow the query.
• get_conversation returns the 100 most recent messages by default, 500 at most.
• Periods are rolling windows ending now — 7d, 30d or 90d. Arbitrary date ranges are not supported. list_leads also accepts no period at all, which means the whole history.
• Requests are rate-limited per account and client. Going over the limit returns a 429 with a Retry-After header — wait and retry.

Support

Questions or issues — contact Elfsight Help Center or your account manager.