Mission · Why this matters
Attach the brain — your agent searches on its own
The line has been live since lesson 1. Today it gets a brain: you connect a chat model to the MCP access, allow your agent exactly two Tools — and then give it its first job in the chat. Afterwards you read in the log what it really did. That is exactly what control looks like when you build it yourself.
What you'll be able to do
- Assemble a KI-Agent in n8n: AI Agent, chat model, and MCP Client Tool.
- Decide via whitelist which Tools your agent is allowed to use at all.
- Read in the execution log which Tool your agent called, and with which parameters.
Three building blocks, one agent
In lesson 1 you fetched the tool catalog by hand. From now on a KI-Agent takes over — built from three n8n building blocks: the AI Agent node is the frame that receives your chat message. A chat model is its brain — it decides which Tool fits the job. And the MCP Client Tool is its hand: the same connection to LinkYourSkill you built in lesson 1, except now the agent uses it instead of you.
Any chat model with tool support works. We show the verified path with Anthropic: the Anthropic Chat Model node with the model Claude Sonnet 5. For that you need your own API key from Anthropic — it pays for the thinking, while your LinkYourSkill Token pays for nothing and stays the same as before.
How to assemble it
- Create a new workflow and add the AI Agent node.
- Attach the Anthropic Chat Model node as the chat model, store your Anthropic API key as a credential, and pick the model Claude Sonnet 5.
- Attach the MCP Client Tool node as a tool.
- Inside it, set Server Transport to
HTTP Streamable. The second option, "Server Sent Events (Deprecated)", is outdated — don't pick it. - In the Endpoint field, enter:
https://app.linkyourskill.ai/mcp - Set Authentication to Bearer Auth and pick the credential from lesson 1 — it's the same Token, and your agent identifies itself exactly the way you did.
- Set Tools to Include to Selected. The selection list now fills with the complete Tool catalog — the node performs the handshake with the platform by itself. When the list appears, you know the connection is right.
- Select exactly two Tools:
search_skillanbieterandget_skillanbieter_ratings. Everything else stays off.
Core idea
The whitelist is the twin of your Auftrags-Limits (order limits):
limits restrict how much may become real — the whitelist restricts what your agent
can even attempt. And precision matters: the catalog contains similar-sounding Tools
like prepare_order (one order) and prepare_orders (many at
once). Your chat model will take any of the allowed ones — which is why in lesson 3
you allow exactly prepare_order, not the plural.
The first chat with your agent
Open the workflow's chat and write to your agent:
Find a cleaner in Münster — show me the best 3.
Your agent then calls search_skillanbieter and receives hits sorted by
distance — providers from Münster come out on top. The "show me the best 3" is
deliberate: without that request, your agent happily presents you the complete list
of hits.
Read what it really did
The answer in the chat is half the truth — the whole truth is in the run's
execution log. There you see the call to search_skillanbieter
including the search query your agent made out of your request, with its own ID per tool
call. The chat model rephrases the query itself — so the log doesn't show your sentence
word for word, but what the agent turned it into. That look is exactly your control:
you don't take your agent's word for it, you read it up.
If it doesn't work
- The tool list under Tools to Include stays empty — then the Endpoint or the Bearer credential is wrong. It's the same handshake as your HTTP Request from lesson 1: check both there first.
- Your agent answers without searching — look at the execution log. If the tool call is missing there, check that the MCP Client Tool is really attached to the AI Agent as a tool and that both Tools are selected.
Your turn
Test yourself — retrieving from memory is what builds real retention.
Knowledge check
What does "Tools to Include → Selected" do on the MCP Client Tool?
"Selected" is your whitelist: your agent sees and uses only the Tools you selected — the twin of your Auftrags-Limits. The selection has nothing to do with cost, and your Token keeps exactly the rights it had before: the whitelist can only narrow rights, never expand them.
