web-search

An LM Studio plugin that gives your local models the ability to search the web and read web pages. Built for agentic, iterative research — not just one search and done.

Features

DuckDuckGo search — scrapes real search results with titles, URLs, and snippets via impit for browser-like fetching
Web page reading — fetches and cleans page content using Readability + Turndown for clean prose, with Jina Reader as fallback
YouTube transcripts — automatically extracts the full transcript when given a YouTube URL instead of scraping the page
PDF support — always routed through Jina for extraction
Site-specific handling — automatic URL rewrites for better content on specific sites:
- Reddit links use old.reddit.com for clean thread access
- arXiv abstract links redirect to the PDF version
- Medium articles route through scribe.rip to bypass the paywall, with fallback to Jina on the original URL if scribe.rip is unavailable
- AMP pages are de-AMPed to the original URL before fetching
- Any URL containing "pdf" is routed through Jina
Result caching — repeated searches within 5 minutes return cached results instantly without hitting the network again

Installation

Install from the LM Studio Hub via "Run in LM Studio" button

Configuration

Setting	Default	Description
Search Results Per Page	5	How many results to return per search (1–10)
Max Content	8000	Maximum characters returned by Visit Website. Increase for long articles, decrease to save context. -1 = auto

Tools

Web Search — searches DuckDuckGo and returns a list of results with titles, URLs, and snippets. Uses Impit for browser-like fetching to evade bot detection. Results for repeated queries are served from an in-memory LRU cache.
Visit Website — fetches a URL and returns cleaned text content. Handles YouTube, PDFs, Reddit, arXiv, Medium, and AMP URLs automatically. Tries Readability + Turndown first for clean, fast results. Falls back to Jina if direct fetch fails. PDFs always use Jina.

Prompt Guidance

The tool descriptions guide the model to follow an iterative research workflow: search → visit one result → visit another → search again → visit one more → visit one more → answer.

This is most useful with smaller models that tend to answer after a single search. Larger models that already do agentic research naturally may require less guidance.

Notes

Search results and content fetching depend on DuckDuckGo and Jina being accessible from your machine — both are free services with no API key required
YouTube transcripts require captions to exist on the video. Auto-generated captions count, but some music videos and very new uploads may not have them
Direct fetch uses Impit, then Readability (same engine as Firefox's reader mode), which strips navigation, sidebars, and other non-content elements. Link URLs are removed, keeping only the text.
Jina is used as fallback when direct fetch fails, and always for PDFs
Search results are cached per session in memory only (up to 100 entries) — cache does not persist across LM Studio restarts

Credits

Originally based on danielsig's DuckDuckGo search and Visit Website plugins for LM Studio.

License

MIT