AnythingLLM MCP web search guide
How to Add Web Search to AnythingLLM with Webstractor MCP
AnythingLLM already ships a Web-Browsing agent skill with a default search option, so an MCP server is not mandatory for basic search. Webstractor is a fit when the same agent workflow runs in more than one MCP host, when source-linked output needs a stable schema, or when search should connect directly to subsequent URL extraction.
Quick answer
Add a streamable server to AnythingLLM’s MCP config.
Open AnythingLLM’s MCP management UI or edit anythingllm_mcp_servers.json in the storage plugins directory. Set type to streamable and URL to https://webstractor.com/mcp. Webstractor complements AnythingLLM’s built-in Web-Browsing skill when you need a reusable provider-neutral tool across deployments.
{
"mcpServers": {
"webstractor": {
"type": "streamable",
"url": "https://webstractor.com/mcp"
}
}
}Before you begin
What you need
- AnythingLLM Desktop or self-hosted Docker with MCP compatibility
- An agent-capable LLM
- Access to the AnythingLLM storage plugins configuration or MCP UI
Step-by-step setup
Connect Webstractor to AnythingLLM
Decide whether Webstractor should replace or complement Web-Browsing
For a clean test, temporarily disable the overlapping Web-Browsing skill or explicitly name Webstractor search_web in prompts. This makes tool selection observable.
Open the MCP configuration
Use the MCP Management UI, or edit anythingllm_mcp_servers.json in the storage plugins directory. Docker users must edit the file inside the mounted AnythingLLM storage path.
Use AnythingLLM’s streamable transport name
AnythingLLM expects type streamable—not streamable-http or streamableHttp. Add this entry inside mcpServers.
{
"mcpServers": {
"webstractor": {
"type": "streamable",
"url": "https://webstractor.com/mcp"
}
}
}Reload and inspect the server
Reload or restart MCP servers in the management UI. Confirm Webstractor is running and that search_web appears in the available tool list.
Start an agentic test chat
Use an agent-enabled chat, or begin with @agent if your selected model needs that directive. Prompt: “Use Webstractor search_web to find five official sources about AnythingLLM MCP compatibility and cite every URL.”
AnythingLLM-specific choice
Built-in browsing is simpler; MCP is more portable
AnythingLLM’s Web-Browsing skill is already configurable and may be sufficient for one workspace. Webstractor becomes useful when you want identical tool names and output behavior in AnythingLLM, local-model desktops, and coding agents.
AnythingLLM’s Intelligent Tool Selection can reduce prompt overhead when many skills and MCP servers are installed. Enable it after confirming the server loads correctly.
Docker
Treat the storage volume as the source of truth
In a self-hosted deployment, the MCP config lives under AnythingLLM storage. Make the edit in the persistent mounted volume, not an ephemeral container layer, then restart or reload MCP servers from the UI.
- Validate JSON before restarting
- Keep the mcpServers wrapper
- Use type: streamable for remote HTTP
- Confirm the server and tools in MCP Management
Troubleshooting
A loaded tool still needs an agent-capable model
AnythingLLM can show a healthy MCP server while the current chat never calls it. Confirm that agent tools are available for the workspace and use @agent when needed.
If a different search tool answers instead, disable the overlap for testing or name Webstractor search_web explicitly.
Available data
What you can extract
- Provider-neutral web result feeds
- Up to 10 ordered results with source URLs
- Markdown or schema-v1 JSON output
- A companion extraction tool for selected results
AI workflows
Where normalized data helps
- Standardize research tools across team workspaces
- Find sources before adding selected pages to a knowledge base
- Run site-scoped documentation discovery
- Build repeatable source-first agent instructions
Boundaries
Public data only
- AnythingLLM uses MCP tools through AI Agents, not ordinary no-tool chat.
- Its built-in Web-Browsing skill may overlap with search_web.
- Do not hard-code paid API credentials in a shared configuration file.
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
AnythingLLM and Webstractor FAQ
Does AnythingLLM already include web search?
Yes. Its Web-Browsing agent skill can search without Webstractor. The MCP option is for portable tools, consistent outputs, and combined discovery-plus-extraction workflows.
What transport value should I use?
Use "streamable" in AnythingLLM. Its configuration name differs from clients that use "streamable-http" or "streamableHttp".
Why do the tools not appear after editing Docker files?
Verify you edited the plugins config in the persistent storage volume, then reload or restart MCP servers in the management UI.
Ready to try it?