n8n MCP Client web search / 2026
How to Build a Web Search Agent in n8n with MCP in 2026
n8n’s MCP Client Tool lets an AI Agent call tools from an external server without wrapping them in separate HTTP Request nodes. A narrow search workflow needs three functional parts: a trigger, a chat model, and an AI Agent whose Tool port is connected to the MCP Client Tool. Selecting only search_web keeps the agent’s tool list small and its behavior easier to predict.
Quick answer
Attach an MCP Client Tool to an n8n AI Agent.
In n8n, connect an MCP Client Tool sub-node to the AI Agent’s Tool port. Choose HTTP Streamable, enter the hosted Webstractor endpoint, select no authentication, and expose only search_web. Then connect a Chat Model and trigger to test the workflow.
MCP Client Tool
Server Transport: HTTP Streamable
Endpoint: https://webstractor.com/mcp
Authentication: None
Tools to Include: Selected → search_webBefore you begin
What you need
- A current n8n release with MCP Client Tool HTTP Streamable transport
- An AI Agent and supported Chat Model
- A manual, chat, webhook, or schedule trigger
Step-by-step setup
Connect Webstractor to n8n
Create the agent workflow
Add a trigger, an AI Agent node, and your Chat Model. Connect the model to the Agent’s Chat Model input. Use Chat Trigger for interactive testing or a production trigger that supplies a query field.
Attach and constrain the MCP Client Tool
Add MCP Client Tool as a sub-node on the Agent’s Tool connector. Select HTTP Streamable, set the endpoint, choose Authentication: None, set Tools to Include to Selected, and choose search_web.
serverTransport: HTTP Streamable
endpoint: https://webstractor.com/mcp
authentication: None
toolsToInclude: Selected
selectedTools:
- search_webTest with inspectable output
Pin a representative query during development. In the agent system message, require search_web for current questions and require URLs in the response.
# Send through Chat Trigger
Find the current n8n MCP Client Tool documentation. Use webstractor search_web and return five sources with URLs.Workflow logic
Map user input to a bounded research task
The model chooses the tool arguments, so define behavior in the Agent system message: use three to five results for narrow questions, pass a site when the request names a domain, and never invent a source URL.
Add a human review or downstream extraction step when the workflow publishes content. Search snippets are useful for selection, but the full source should support material claims.
Use the webstractor search_web tool for questions that require current public sources.
Return source URLs with every researched answer.
Prefer 5 results and restrict by site when the user names a domain.
Do not treat a result snippet as full-page evidence.Troubleshooting
Check transport before credentials
Use HTTP Streamable for the hosted /mcp endpoint. Webstractor public access does not need a credential, so select None rather than creating a placeholder header.
If the tool connects but is absent from the agent, verify the sub-node is connected to the Tool port and that search_web is included by the filter. For expressions in sub-nodes, remember n8n resolves them against the first input item.
Available data
What you can extract
- Ordered public search results
- Titles, URLs, and concise snippets
- Search limits and locale controls
- Optional site and PDF constraints
AI workflows
Where normalized data helps
- Research before drafting a support response
- Find sources before a content workflow
- Discover public documentation from a ticket
- Collect candidate links for a human review step
Boundaries
Public data only
- The MCP Client Tool is an AI sub-node, not a standalone workflow trigger.
- Search tool calls consume the normal workflow execution context.
- Search results should be read before an agent treats snippets as evidence.
webstractor.com does not bypass CAPTCHAs, login walls, paywalls, access controls, or regional restrictions. Review the source’s terms and applicable law before collecting or reusing data.
Common questions
n8n and Webstractor FAQ
Is MCP Client Tool the same as MCP Server Trigger?
No. MCP Client Tool lets an n8n agent call an external server. MCP Server Trigger exposes n8n tools to external MCP clients.
Which transport should I select?
Choose HTTP Streamable for https://webstractor.com/mcp.
Should I expose every Webstractor tool?
Not for a search-only agent. Select search_web to minimize tool ambiguity, then add extraction separately when needed.
Ready to try it?