Langflow MCP search flow / 2026
How to Add Web Search to a Langflow Agent with MCP in 2026
Langflow makes the MCP connection visible on the canvas. That is useful for search because the MCP Tools component becomes an explicit dependency of the Agent rather than an invisible capability buried in a prompt.
Quick answer
Register Webstractor, then connect MCP Tools to the Agent.
In Langflow’s MCP sidebar, add the hosted Webstractor URL, wait for actions to load, drag its MCP Tools component onto the canvas, enable Tool Mode, and connect Toolset to the Agent’s Tools port.
{
"mcpServers": {
"webstractor": { "url": "https://webstractor.com/mcp" }
}
}Before you begin
What you need
- A current Langflow project
- A Simple Agent flow or equivalent Agent component
- A configured model provider
- Permission to add MCP servers
Step-by-step setup
Connect Webstractor to Langflow
Register the remote MCP server
Open MCP in the left sidebar, choose Add MCP Server, select the JSON option, and add Webstractor by URL.
{
"mcpServers": {
"webstractor": {
"url": "https://webstractor.com/mcp"
}
}
}Wait for action discovery
Click Add Server and wait until the actions list populates. This confirms that Langflow completed MCP initialization and tool discovery.
Wire MCP Tools into the flow
Drag the webstractor server from the MCP sidebar to the canvas. Enable Tool Mode on the component and connect its Toolset output to the Agent component’s Tools input.
Set a search discipline in the Agent
Tell the model to search only when it needs public information and to keep result URLs in the final response.
instructions: |
Use search_web for public source discovery.
Start with 5 results and one concise query.
Preserve source URLs in the answer.
Read selected pages separately when snippets are insufficient.Canvas architecture
Make search an inspectable tool dependency
Keep Chat Input connected to Agent input, MCP Tools connected to Agent tools, and Agent output connected to Chat Output. This simple topology exposes the point at which external context enters the answer.
For multi-agent projects, connect search only to the researcher agent. A writer or formatter agent can receive curated context without gaining another path to call the web.
Verification
Use Playground and read the tool trace
Ask for three sources about a narrow topic and verify that search_web receives the query and a bounded limit. The final response should include URLs returned by the tool.
Then ask a follow-up about one result. If the answer needs page detail, the agent should say it needs to read the page or use a separately enabled extraction tool rather than overclaim from the snippet.
Troubleshooting
Fix connection and selection problems separately
An empty Actions list points to registration, URL, or server-network access. A populated list with no tool calls points to the model or its instructions. Keeping those failure classes separate saves time.
If multiple MCP servers expose similarly described search tools, remove unused servers from this flow or give the Agent an explicit preference.
Available data
What you can extract
- Up to 10 ordered public web results
- Source titles, URLs, and snippets
- Language, country, and hostname restrictions
- A reusable MCP Tools component on the flow canvas
AI workflows
Where normalized data helps
- Visual research-agent prototypes
- Website-scoped documentation assistants
- Search-before-extract Langflow pipelines
- Source-linked answer generation in Playground
Boundaries
Public data only
- An MCP Tools component must be connected to the Agent Tools port.
- Search snippets are not full-page evidence.
- The deployed Langflow runtime, not only the browser, needs network access to Webstractor.
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
Langflow and Webstractor FAQ
Do I need one MCP Tools component per server?
Yes. Langflow documents one MCP Tools component for each connected MCP server.
Can I reuse the connection in another flow?
The registered server appears in the MCP sidebar and can be added as a component where needed.
Why does the agent answer without searching?
Confirm the model supports tools, the Toolset port is connected, and the instructions clearly state when search_web is required.
Ready to try it?