The knowledge base as an MCP connector.

One URL. This is a remote MCP server over Streamable HTTP — no installation, no local process, no API key.

Responses come back as text/event-stream, even for a single call. Any common MCP client handles that for you; if you work with raw curl, you'll see one event: message line per response.

Claude

Works in Claude on the web and in the desktop app.

1. 1Open Settings → Connectors.
2. 2Click Add custom connector.
3. 3Give it a name and paste the endpoint above as the URL.
4. 4Confirm. The four tools show up immediately — there's no sign-in step, because there's no auth.

On Team and Enterprise plans only a workspace admin can usually add connectors. If you're stuck on step 2, that's almost always why.

Claude Code

In the CLI it's a single command:

claude mcp add --transport http kennisbank https://kennisbank.gentrepreneur.be/api/mcp

ChatGPT

ChatGPT supports custom MCP connectors on the paid plans. The route runs through Settings → Connectors → Advanced, where you enable developer mode, then Add connector with the same URL. Pick No authentication when asked.

OpenAI moves these screens around regularly and availability differs per plan. If you can't find the menu, the connector still works — through the API or one of the clients below.

Cursor, VS Code, Zed and other clients

These take the same block of JSON. The file name differs per tool (mcp.json, .mcp.json, settings.json), the contents don't:

{
  "mcpServers": {
    "kennisbank": {
      "type": "http",
      "url": "https://kennisbank.gentrepreneur.be/api/mcp"
    }
  }
}

Every tool takes language ("nl" or "en", default nl) and returns JSON with absolute URLs to the public article pages.

Without an SDK, plain JSON-RPC 2.0 over POST works. Important: the Accept header must contain both types, or the server refuses the request.

curl -X POST https://kennisbank.gentrepreneur.be/api/mcp \
  -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://kennisbank.gentrepreneur.be/api/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tools/call",
    "params": {
      "name": "search_knowledge_base",
      "arguments": { "query": "btw-nummer aanvragen", "language": "nl", "limit": 3 }
    }
  }'

The result comes back as a single data: line containing a content array; the first text block holds the tool's JSON.

Prefer it without MCP

If you don't have an MCP client, the same four tools are available as a plain JSON API. One path per tool, GET with a query string or POST with a JSON body:

curl 'https://kennisbank.gentrepreneur.be/api/agent/search_knowledge_base?query=startlening&limit=3'

curl -X POST https://kennisbank.gentrepreneur.be/api/agent/ask_knowledge_base \
  -H 'Content-Type: application/json' \
  -d '{"question":"Hoe vraag ik een startlening aan?","language":"nl"}'

Server-to-server only. This path sends no CORS headers, so it won't work straight from browser JavaScript. Call it from your own backend, or use the MCP endpoint — that one is open to cross-origin clients.

Rate limits apply per IP address, on a sliding one-minute window. They're separate from our own chatbot's budget, so normal agent use won't come near them.

Exceed a limit and you don't get a crash but a tool result with an error message in it. Waiting a moment is enough.

• Published content only. Drafts and internal articles are not in this channel and won't be added to it.
• Credit the source. Every result carries the article's public URL. Link to it when you pass our content on to an end user.
• The content is alive. Articles are updated continuously; cache briefly at most and don't assume a stable slug.
• No SLA. This is a free, public service. If you're building something business-critical on it, let us know — then we'll look at arrangements together.

Questions, a use case we should know about, or a tool you're missing? Email Luc Peeters — luc.peeters@gentrepreneur.be.

/llms.txt — Machine-readable entry point for agents