mcp-server-webcrawl

Advanced search and retrieval for web crawler data. With mcp-server-webcrawl, your AI client filters and analyzes web content under your direction or autonomously. The server includes a fulltext search interface with boolean support, and resource filtering by type, HTTP status, and more.

mcp-server-webcrawl provides the LLM a complete menu with which to search, and works with a variety of web crawlers:

mcp-server-webcrawl is free and open source, and requires Claude Desktop and Python (>=3.10). It is installed on the command line, via pip install:

pip install mcp-server-webcrawl

For step-by-step MCP server setup, refer to the Setup Guides.

Features

• Claude Desktop ready
• Multi-crawler compatible
• Filter by type, status, and more
• Boolean search support
• Support for Markdown and snippets
• Roll your own website knowledgebase

Prompt Routines

mcp-server-webcrawl provides the toolkit necessary to search web crawl data freestyle, figuring it out as you go, reacting to each query. This is what it was designed for.

It is also capable of running routines (as prompts). You can write these yourself, or use the ones provided. These prompts are copy and paste, and used as raw Markdown. They are enabled by the advanced search provided to the LLM; queries and logic can be embedded in a procedural set of instructions, or even an input loop as is the case with Gopher Service.

If you want to shortcut the site selection (one less query), paste the markdown and in the same request, type "run pasted for [site name or URL]." It will figure it out. When pasted without additional context, you should be prompted to select from a list of crawled sites.

Boolean Search Syntax

The query engine supports field-specific (field: value) searches and complex boolean expressions. Fulltext is supported as a combination of the url, content, and headers fields.

While the API interface is designed to be consumed by the LLM directly, it can be helpful to familiarize yourself with the search syntax. Searches generated by the LLM are inspectable, but generally collapsed in the UI. If you need to see the query, expand the MCP collapsible.

Example Queries

Field Search Definitions

Field search provides search precision, allowing you to specify which columns of the search index to filter. Rather than searching the entire content, you can restrict your query to specific attributes like URLs, headers, or content body. This approach improves efficiency when looking for specific attributes or patterns within crawl data.

Field Content

A subset of fields can be independently requested with results, while core fields are always on. Use of headers and content can consume tokens quickly. Use judiciously, or use extras to crunch more results into the context window. Fields are a top level argument, independent of any field searching taking place in the query.

Content Types

Crawls contain resource types beyond HTML pages. The type: field search allows filtering by broad content type groups, particularly useful when filtering images without complex extension queries. For example, you might search for type: html NOT content: login to find pages without "login," or type: img to analyze image resources. The table below lists all supported content types in the search system.

Extras

The extras parameter provides additional processing options, transforming HTTP data (markdown, snippets, regex, xpath), or connecting the LLM to external data (thumbnails). These options can be combined as needed to achieve the desired result format.

Extras provide a means of producing token-efficient HTTP content responses. Markdown produces roughly 1/3 the bytes of the source HTML, snippets are generally 500 or so bytes per result, and XPath can be as specific or broad as you choose. The more focused your requests, the more results you can fit into your LLM session.

The idea, of course, is that the LLM takes care of this for you. If you notice your LLM developing an affinity to the "content" field (full HTML), a nudge in chat to budget tokens using the extras feature should be all that is needed.

Interactive Mode

No AI, just classic Boolean search of your web-archives in a terminal.

mcp-server-webcrawl can double as a terminal search for your web archives. You can run it against your local archives, but it gets more interesting when you realize you can ssh into any remote host and view archives sitting on that host. No downloads, syncs, multifactor logins, or other common drudgery required. With interactive mode, you can be in and searching a crawl sitting on a remote server in no time at all.

Launch with --crawler and --datasrc to search immediately, or setup datasrc and crawler in-app.

mcp-server-webcrawl --crawler wget --datasrc /path/to/datasrc --interactive
# or manually enter crawler and datasrc in the UI
mcp-server-webcrawl --interactive

Interactive mode is a way to search through tranches of crawled data, whenever, whereever... in a terminal.