LM Studio Web Search Plugin

A small LM Studio plugin that gives a local model three tools — web_search, visit_website, and image_search — backed by DuckDuckGo's public HTML interface. No API keys, no accounts, no third-party search backend to configure.

This is a fork of danielsig/lms-plugin-duckduckgo. The main divergence is a dedicated visit_website tool plus tool descriptions tuned to stop the model from "searching" full URLs instead of fetching them.

Features

• Three-tool surface the LLM understands: web_search (links only), visit_website (full page text via Mozilla Readability), and image_search (downloads the actual files to the plugin's working directory).
• Search → fetch chaining baked into the tool docstrings. web_search deliberately returns no page bodies and tells the model to call visit_website next; visit_website tells it to cite the source URL.
• URL guardrail. Passing a URL to web_search or image_search returns an error string that names the right tool, instead of silently producing a junk search.
• Current-date injection in the web_search response, so the model can resolve "yesterday" / "this week" against the real date instead of its training cutoff.
• DuckDuckGo HTML scraping, not the Instant Answer API — you get the same results a logged-out user sees in a browser. Requests rotate a pool of real-world User-Agents and a shared 2 s rate limiter.
• Image search downloads to disk. Image URLs are fetched in parallel into the plugin's working directory and the local file paths are returned, so the model can reference them as local files.

Requirements

• LM Studio with plugin support (Node runner).
• Node.js to build from source. The runtime depends on @lmstudio/sdk, @mozilla/readability, jsdom, and zod.

Installation

One-click

lmstudio.ai/richardbenedikt/web-search → Run in LM Studio.

From the CLI

lms get richardbenedikt/web-search       # install the published plugin
lms clone richardbenedikt/web-search     # clone for local development

Configuration

Both fields live in the plugin sidebar in LM Studio.

Field	Type	Default	Effect
pageSize	int (0–10)	0 (auto)	Number of results per search call. 0 defers to the per-call argument or the built-in default of 5.
safeSearch	enum	auto	strict / moderate / off map to DuckDuckGo's kp parameter. auto defers to the per-call argument, falling back to moderate.

The "auto" sentinel exists so the model can override the sidebar setting per call when the user explicitly asks for it.

How the model uses it

The first sentence of each tool's description is what the LLM sees and gates on. They are deliberately blunt and chained:

• web_search — "Step 1: Find relevant URLs. Returns a list of links but NO content. If results are relevant, you MUST use 'visit_website' to read the page. If not, refine the query and search again."
• visit_website — "Step 2: Read the content of a specific URL found via 'web_search'. You can call this multiple times to verify information across different sources."
• image_search — "Search for images on DuckDuckGo. Returns a list of image URLs."

web_search returns a { candidates, current_date, status: "incomplete", system_instruction } object whose system_instruction reminds the model that titles aren't enough and that source URLs must be cited. visit_website appends the same instruction as a trailing SYSTEM INSTRUCTION: line. These strings are part of the API surface — edit them with care.

Example flow

1. User: "What changed in the latest LM Studio release?"
2. Model: calls web_search({ query: "LM Studio latest release notes" }).
3. Model: picks the most relevant link and calls visit_website({ url: "..." }) (often two or three of them).
4. Model: answers with the source URLs cited inline.

Development

git clone https://github.com/richardbenedikt/lms-plugin-websearch.git
cd lms-plugin-websearch
npm install

npm run build   # tsc → dist/
npm run dev     # lms dev — live-loads the plugin into a running LM Studio instance
npm run push    # lms push — publishes a new revision (bump manifest.json `revision` first)

There is no test runner or linter wired up; lms dev is the iteration loop.

Source layout:

src/
├── index.ts            # plugin entry — registers config + tools provider
├── toolsProvider.ts    # builds the three tools, shares one rate limiter
├── config.ts           # sidebar fields (pageSize, safeSearch)
├── utils.ts            # spoofed headers, rate limiter, safe-search mapping
└── tools/
    ├── webSearch.ts
    ├── visitWebsite.ts
    └── imageSearch.ts

Add a tool by writing a createXTool(ctl, waitIfNeeded) factory in src/tools/ and pushing the result into the array in toolsProvider.ts. Always await waitIfNeeded() before any outbound HTTP call — the limiter is shared across tools and DuckDuckGo will start returning empty pages or CAPTCHAs across all three if any tool skips it.

Privacy

visit_website and image_search make direct network requests to the target URLs and image CDNs from the machine running LM Studio. Those servers see your IP. There is no proxy or anonymization layer.

Limitations

• DuckDuckGo HTML scraping. If DDG changes its markup, anti-bot rules, or VQD-token flow, all three tools break before anything else does.
• visit_website parses with JSDOM + @mozilla/readability and does not execute JavaScript. Pages that render entirely client-side return little useful text. Page bodies are truncated to 20 000 characters.
• image_search filters to .jpg / .png / .gif / .jpeg only and writes files into ctl.getWorkingDirectory() named <timestamp>-<index>.<ext>. Old images are not cleaned up.
• The shared 2 s rate limit caps throughput at roughly one request every two seconds across all tools.

License

MIT. Fork of danielsig/lms-plugin-duckduckgo.