MCP server for your knowledge base

What a knowledge base MCP server does

The Model Context Protocol (MCP) is an open standard that lets AI assistants call external tools and data sources in a uniform way. Instead of copying articles into a prompt or hoping a model was trained on your docs, the assistant asks your knowledge base directly: it searches, reads the relevant article, and answers with a link to the source.

For a help center this closes a real gap. Your customers and your own team increasingly ask questions inside AI assistants rather than on your website. An MCP endpoint makes your current, published documentation the thing those assistants read, so answers reflect today's product and not a months-old crawl.

Your endpoint: one per portal, live by default

Each portal is exposed automatically. There is nothing to install, no API key to issue and no server to host. If you run several branded portals, each one has its own endpoint and only ever returns that portal's content.

• URL: https://be.multibase.io/api/mcp/<portal-slug>, also served on the portal's own host at /mcp (including your custom domain).
• Transport: Streamable HTTP, stateless, JSON responses.
• Compatible with any MCP client that supports remote Streamable HTTP servers, for example Claude, Cursor and VS Code.
• Articles are returned as Markdown with their canonical URL, so assistants can cite and link back to your portal.

Connect an MCP client

Most clients accept a remote server as a name plus a URL. The snippet below uses the common mcpServers shape. The exact configuration keys differ per client: VS Code, for example, uses a servers object with "type": "http", and Claude lets you add a remote server as a custom connector by pasting the URL. Check your client's documentation for the precise format.

{
  "mcpServers": {
    "acme-help": {
      "url": "https://be.multibase.io/api/mcp/acme"
    }
  }
}

Test it with curl

Because the endpoint is stateless JSON-RPC over HTTP, you can inspect it without any client. Send a POST with both content-type and accept headers set as shown. The first call lists the available tools; the second runs a search.

curl -X POST https://be.multibase.io/api/mcp/acme \
  -H "content-type: application/json" \
  -H "accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

curl -X POST https://be.multibase.io/api/mcp/acme \
  -H "content-type: application/json" \
  -H "accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call",
       "params":{"name":"search_kb","arguments":{"query":"reset password","limit":5}}}'

Tools reference

The locale argument lets an assistant request a specific language when the article has been published in it, which matters for portals that ship content in several of the 9 supported languages.

Security and scope

• Public, published content only. Drafts, unpublished locales and articles restricted to reader groups are never exposed.
• Read-only. No tool can create, edit or delete anything.
• Rate-limited per endpoint to protect your portal and keep usage predictable.
• Tenant-isolated. A portal's endpoint returns only that portal's articles.
• Visible in analytics. MCP searches and article reads appear with the source "MCP", next to web and AI chat traffic, so you can see what assistants ask about.

Use cases

• Support agents working in Claude or Cursor answer tickets from your actual docs, with links they can paste to customers.
• Internal copilots and engineering assistants look up product behaviour without anyone maintaining a separate copy of the docs.
• Your customers' own AI agents connect to your public help center and resolve how-to questions without opening a ticket.
• Failed MCP searches feed the same analytics as reader searches, so gaps surfaced by assistants become article ideas.

What is not available yet

Authenticated access is not available today. There is no OAuth flow and no way for an MCP client to read reader-group or other private content; the endpoint covers public content only. If your main use case is giving AI agents access to internal or permission-scoped documentation, Multibase's MCP server does not cover that yet, and a tool built around authenticated access will serve you better for now.