No browser, no Playwright, no API keys — just Node 22's built-in fetch and a small HTML→text extractor.

Tools exposed to the model

All search responses include returned, dropped_as_noise, from_cache, and per_engine_counts for transparency. When the top result score is very low (< 0.35) or zero results survive filtering, an additional note field tells the model to rephrase or pick a more specific tool.

Result ranking & noise reduction

Every web_search / web_search_full response is post-processed:

• Junk drop — empty titles, search-engine redirect wrappers (duckduckgo.com/l/?, google.com/url?, etc.), non-http URLs.
• URL normalization — lowercase host, strip www., strip #fragment, drop utm_* / fbclid / ref / source / gclid tracking params, strip trailing /.
• Dedupe — by normalized URL, keeping the higher-scoring instance.
• Score = 0.45 · titleQueryOverlap + 0.30 · hostTrust + 0.20 · snippetQueryOverlap + 0.05 · snippetLengthBonus.
  • Host trust is a built-in map: wikipedia/arxiv ≈ 0.9, .gov 0.85, .edu / github / stackoverflow ≈ 0.8, reddit ≈ 0.55, pinterest ≈ 0.2, unknown 0.5.
• Top-K — sorted by score, sliced to topK (default 8).
• Snippet trim — cap each snippet to snippetMaxChars (default 240).

Turn on includeScoreBreakdown (per-chat setting / MULTI_SEARCH_SCORE_BREAKDOWN=1) to see per-result { host_trust, title_overlap, snippet_overlap, snippet_length_bonus }.

In-process cache

A tiny LRU+TTL cache (100 search entries, 50 web_search_full entries) deduplicates repeat calls within the same plugin session. Default TTL is 5 minutes. Set cacheTtlSec to 0 (or MULTI_SEARCH_CACHE_TTL=0) to disable. Responses include from_cache: true when served from cache.

Supported engines (14)

duckduckgo, brave, bing, qwant, ecosia, startpage, metager, wikipedia, arxiv, reddit, stackoverflow, github, google-scholar, devdocs

Most reliable: duckduckgo, brave, bing, wikipedia, arxiv, reddit, stackoverflow, github. Scraped engines (ecosia, startpage, etc.) can occasionally be rate-limited — the multi-engine mode tolerates this and just keeps whatever returned successfully.

Install

git clone https://github.com/potkolainen/LM-Multi-Online-Search-plugin.git
cd "Multi search"
npm install
npm run build

Requires Node 22+ (uses the built-in global fetch / AbortController).

Use as a standalone MCP server

After npm run build, point your MCP host at dist/mcp.js. Example mcp.json (LM Studio, Claude Desktop, etc.):

{
  "mcpServers": {
    "multi-search": {
      "command": "node",
      "args": ["/absolute/path/to/Multi search/dist/mcp.js"],
      "cwd": "/absolute/path/to/Multi search",
      "env": {
        "MULTI_SEARCH_MODE": "multi",
        "MULTI_SEARCH_MULTI_ENGINES": "duckduckgo,brave,bing,wikipedia",
        "MULTI_SEARCH_FETCH_TOP_N": "3",
        "MULTI_SEARCH_MAX_CONTENT": "10000"
      }
    }
  }
}

Environment variables

MCP has no settings UI, so the standalone server is configured purely via env vars:

Quick smoke test

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"x","version":"0"}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
| node dist/mcp.js

Should print two JSON-RPC responses on stdout and multi-search MCP server ready on stdio on stderr.

Use as an LM Studio plugin

npm install
lms dev          # hot-reload while developing
# or
lms push         # install into LM Studio

When loaded as a plugin you get a per-chat config panel with these fields:

And global plugin settings (apply across all chats):

How multi-engine mode works

• Fan out the same query to every checked engine in parallel (Promise.all).
• Round-robin interleave: take the 1st result from each engine, then 2nd from each, then 3rd…
• Dedupe by URL (ignoring ?query and #fragment).
• Tag each result with its source engine so the model knows where it came from.
• Any engine that errors or times out returns [] — the rest still come through.

This gives broader topical coverage than a single engine and degrades gracefully when one engine breaks.

Page extraction

fetch_page and web_search_full use a small dependency-free extractor that:

• strips <script>, <style>, <noscript>, <svg>, <iframe>, <form>, <nav>, <aside>, <footer>, <header>,
• prefers <article> → <main> → largest <div> by visible text,
• decodes HTML entities,
• converts block-level tags to newlines (so <p>foo</p><p>bar</p> doesn't collapse to foobar),
• collapses whitespace,
• truncates to maxContentLength.

It is not a full Readability re-implementation — for pathological pages (heavy SPA, lazy-loaded content, Cloudflare challenges) it will return less than a browser-based scraper would. The trade-off is zero native deps and instant cold start.

Project layout

src/
  engines.ts          # 14 search engines + shared helpers (fetch with timeout, etc.)
  extract.ts          # HTML → main-text extractor (with query-aware truncation)
  ranking.ts          # URL normalization, dedupe, host-trust scoring, top-K filter
  cache.ts            # Tiny in-process LRU + TTL cache
  weather.ts          # wttr.in client for the get_weather tool
  configSchematics.ts # LM Studio plugin per-chat + global config schema
  toolsProvider.ts    # LM Studio plugin tool registration
  index.ts            # LM Studio plugin entry point
  mcp.ts              # Standalone MCP stdio server entry point

Both distributions share engines.ts and extract.ts, so any new engine becomes available in both modes at once.

Scripts

Adding a new engine

• Implement an EngineFn in src/engines.ts:

• Register it in the ENGINES map.
• Add a pretty label to ENGINE_LABELS in src/configSchematics.ts.
• (Optional) Add it to DEFAULT_MULTI if you want it ticked by default.
• npm run build. It now appears in both the plugin's checkbox list and the MCP server's list_search_engines output.

License

MIT

No browser, no Playwright, no API keys — just Node 22's built-in fetch and a small HTML→text extractor.

Tools exposed to the model

All search responses include returned, dropped_as_noise, from_cache, and per_engine_counts for transparency. When the top result score is very low (< 0.35) or zero results survive filtering, an additional note field tells the model to rephrase or pick a more specific tool.

Result ranking & noise reduction

Every web_search / web_search_full response is post-processed:

• Junk drop — empty titles, search-engine redirect wrappers (duckduckgo.com/l/?, google.com/url?, etc.), non-http URLs.
• URL normalization — lowercase host, strip www., strip #fragment, drop utm_* / fbclid / ref / source / gclid tracking params, strip trailing /.
• Dedupe — by normalized URL, keeping the higher-scoring instance.
• Score = 0.45 · titleQueryOverlap + 0.30 · hostTrust + 0.20 · snippetQueryOverlap + 0.05 · snippetLengthBonus.
  • Host trust is a built-in map: wikipedia/arxiv ≈ 0.9, .gov 0.85, .edu / github / stackoverflow ≈ 0.8, reddit ≈ 0.55, pinterest ≈ 0.2, unknown 0.5.
• Top-K — sorted by score, sliced to topK (default 8).
• Snippet trim — cap each snippet to snippetMaxChars (default 240).

Turn on includeScoreBreakdown (per-chat setting / MULTI_SEARCH_SCORE_BREAKDOWN=1) to see per-result { host_trust, title_overlap, snippet_overlap, snippet_length_bonus }.

In-process cache

A tiny LRU+TTL cache (100 search entries, 50 web_search_full entries) deduplicates repeat calls within the same plugin session. Default TTL is 5 minutes. Set cacheTtlSec to 0 (or MULTI_SEARCH_CACHE_TTL=0) to disable. Responses include from_cache: true when served from cache.

Supported engines (14)

duckduckgo, brave, bing, qwant, ecosia, startpage, metager, wikipedia, arxiv, reddit, stackoverflow, github, google-scholar, devdocs

Most reliable: duckduckgo, brave, bing, wikipedia, arxiv, reddit, stackoverflow, github. Scraped engines (ecosia, startpage, etc.) can occasionally be rate-limited — the multi-engine mode tolerates this and just keeps whatever returned successfully.

Install

git clone https://github.com/potkolainen/LM-Multi-Online-Search-plugin.git
cd "Multi search"
npm install
npm run build

Requires Node 22+ (uses the built-in global fetch / AbortController).

Use as a standalone MCP server

After npm run build, point your MCP host at dist/mcp.js. Example mcp.json (LM Studio, Claude Desktop, etc.):

{
  "mcpServers": {
    "multi-search": {
      "command": "node",
      "args": ["/absolute/path/to/Multi search/dist/mcp.js"],
      "cwd": "/absolute/path/to/Multi search",
      "env": {
        "MULTI_SEARCH_MODE": "multi",
        "MULTI_SEARCH_MULTI_ENGINES": "duckduckgo,brave,bing,wikipedia",
        "MULTI_SEARCH_FETCH_TOP_N": "3",
        "MULTI_SEARCH_MAX_CONTENT": "10000"
      }
    }
  }
}

Environment variables

MCP has no settings UI, so the standalone server is configured purely via env vars:

Quick smoke test

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"x","version":"0"}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
| node dist/mcp.js

Should print two JSON-RPC responses on stdout and multi-search MCP server ready on stdio on stderr.

Use as an LM Studio plugin

npm install
lms dev          # hot-reload while developing
# or
lms push         # install into LM Studio

When loaded as a plugin you get a per-chat config panel with these fields:

And global plugin settings (apply across all chats):

How multi-engine mode works

• Fan out the same query to every checked engine in parallel (Promise.all).
• Round-robin interleave: take the 1st result from each engine, then 2nd from each, then 3rd…
• Dedupe by URL (ignoring ?query and #fragment).
• Tag each result with its source engine so the model knows where it came from.
• Any engine that errors or times out returns [] — the rest still come through.

This gives broader topical coverage than a single engine and degrades gracefully when one engine breaks.

Page extraction

fetch_page and web_search_full use a small dependency-free extractor that:

• strips <script>, <style>, <noscript>, <svg>, <iframe>, <form>, <nav>, <aside>, <footer>, <header>,
• prefers <article> → <main> → largest <div> by visible text,
• decodes HTML entities,
• converts block-level tags to newlines (so <p>foo</p><p>bar</p> doesn't collapse to foobar),
• collapses whitespace,
• truncates to maxContentLength.

It is not a full Readability re-implementation — for pathological pages (heavy SPA, lazy-loaded content, Cloudflare challenges) it will return less than a browser-based scraper would. The trade-off is zero native deps and instant cold start.

Project layout

src/
  engines.ts          # 14 search engines + shared helpers (fetch with timeout, etc.)
  extract.ts          # HTML → main-text extractor (with query-aware truncation)
  ranking.ts          # URL normalization, dedupe, host-trust scoring, top-K filter
  cache.ts            # Tiny in-process LRU + TTL cache
  weather.ts          # wttr.in client for the get_weather tool
  configSchematics.ts # LM Studio plugin per-chat + global config schema
  toolsProvider.ts    # LM Studio plugin tool registration
  index.ts            # LM Studio plugin entry point
  mcp.ts              # Standalone MCP stdio server entry point

Both distributions share engines.ts and extract.ts, so any new engine becomes available in both modes at once.

Scripts

Adding a new engine

• Implement an EngineFn in src/engines.ts:

• Register it in the ENGINES map.
• Add a pretty label to ENGINE_LABELS in src/configSchematics.ts.
• (Optional) Add it to DEFAULT_MULTI if you want it ticked by default.
• npm run build. It now appears in both the plugin's checkbox list and the MCP server's list_search_engines output.

License

MIT