n8n MCP Server and MCP Client in two separate workflows: a small research agent I built to keep the AI’s reasoning apart from the tools it calls.
Workflow 1: MCP_Server
- MCP Server Trigger
- One tool on it: Hacker News → Get many items (getAll)
Workflow 2: MCP_Client
- When chat message received → AI Agent
- Chat model: Google Gemini
- Tool: MCP Client, pointed at the server’s URL
Why split it
The agent only decides what it needs. The server decides how each tool actually runs. So:
- a new tool is one more node on the MCP Server Trigger, and the agent workflow doesn’t change
- Gemini can be swapped for Claude or GPT in the client without touching the tools
Tips if you build one
- The MCP Client needs the server’s Production URL, and the server workflow must be active for that URL to answer. The Test URL only works while the editor is listening.
- Give each tool a clear description. It’s what the agent reads when it decides which tool to call.
- Turn on authentication in the MCP Server Trigger before the server is used outside testing.
What would you put behind an MCP server first?
Something read-only, like your HN tool but pointed at your own data. For example a search on your tickets or CRM that can only get records and never update them.
Your split helps here too, I guess. If the agent only decides what and the server decides how, the server is also where you decide what each caller can touch, and that stays the same when you swap Gemini for Claude. The auth on the MCP Server Trigger is where that starts. I wouldn’t put a write tool behind it until you know who is calling and can limit what that caller sees.
One thing I’d test with this separation is how the agent handles tool failures.
For example, an empty Hacker News result, an API timeout, and a rate-limit error shouldn’t all look like an empty list to the client
I’d give each tool a consistent response structure with a status, source, retrieval timestamp, and an explicit error state where needed
That way, the agent can distinguish “nothing found” from “couldn’t retrieve the data” instead of producing a confident answer from an incomplete result.
The model may change, but that reliability contract can stay the same.