Web Search for LM Studio LLMS

Web Search for LM Studio LLMS

An open-source Model Context Protocol server that gives LM Studio and other MCP clients tools for fresh web search, source-page reading, news, weather, and selected public data. The package and executable remain mcp-internet. It uses the local machine clock and includes a compact retrieval timestamp in research results, so a model can reason about “today” without a separate clock call on every question.

Search results are evidence leads, not guaranteed truth. The server returns source URLs and page content when a site allows access; the model should cite those sources and say when evidence is missing or contradictory. The LM Studio prompt sets a short default answer (80 words; 50 for broad “tell me about” questions), requires citations, and limits claims to returned evidence. Replace the prompt saved in LM Studio and start a new chat for the change to apply.

Features

• Fresh metasearch for each search request; search answers are not served from a local cache.
• smart_search reads up to two sources by default for a faster evidence-backed answer; deep_search fetches a broader set. Both return source URLs, page content when available, publication metadata, and retrieval time.
• Time-sensitive queries receive the server's current local date; timeless and explicitly historical queries are left intact. Search results include a server retrieval timestamp with UTC offset. Explicit recency requests can use search-engine date filters; upcoming schedules avoid narrow date filters so older announcements remain discoverable.
• Source results include readable page text alongside extracted JSON-LD, OpenGraph metadata, publication/update dates, and HTML tables when available, with additional parsers for selected domains.
• Thirteen focused MCP tools, exposed over stdio for local clients.
• No search API key is required by default. Public search engines and websites can rate-limit or change their pages.
• URL reads reject non-public destinations and pin each validated hostname to its resolved public address for the connection, including after redirects.

The goal is source-grounded research for local models: quick search when you need leads, and source reading when you need evidence. Search engines provide candidates; the server labels what it fetched and what it could not verify.

Tools

Example

Ask LM Studio: “What changed in the latest Python release?” The recommended prompt calls smart_search, which returns search leads and readable page evidence when available. A result is shaped like this (field names shown; values depend on the live search):

Search attempted: <retrieval timestamp>
1. <result title>
URL: <source URL>
Snippet: <search-engine snippet>

## Read source pages
## <page title>
URL: <source URL>
### Extracted metadata
- Page published: <date, when present>
### Page text
<bounded extracted text, when reading succeeded>

Use deep_search when you need more source pages. A snippet without a matching extracted page is only a lead, not verified page content.

Quick start

Requirements: Python 3.10 or later, Git, and uv. LM Studio does not install these prerequisites when you add an MCP entry. This project currently recommends running from a local source checkout; no PyPI installation command is advertised.

git clone https://github.com/Sonesh-2202/mcp-internet.git
cd mcp-internet
uv sync

For PDF support:

uv sync --extra pdf

To upgrade an existing checkout, run git pull and then uv sync. To uninstall, remove the mcp-internet entry from LM Studio's mcp.json and delete the checkout. If you enabled PDF support, uv sync without --extra pdf removes that optional dependency from the project environment.

Configure LM Studio

LM Studio 0.3.17 or newer supports an “Add to LM Studio” deeplink. Add Web Search for LM Studio LLMS adds a template entry. After adding it, edit both path values to your checkout path. The deeplink only configures LM Studio; Python, uv, and the project dependencies must already be installed as described above.

You can also add the following entry to LM Studio's mcp.json, replacing the example path with the absolute path to this checkout. Keep the existing outer mcpServers object and other entries if the file already has other servers.

{
  "mcpServers": {
    "mcp-internet": {
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "C:/path/to/mcp-internet",
        "mcp-internet"
      ],
      "cwd": "C:/path/to/mcp-internet"
    }
  }
}

Restart or refresh MCP servers in LM Studio. For platform-specific examples, see the multi-platform guide.

Search behavior and accuracy

What a research response contains

For a factual question, smart_search returns a compact set of result titles, URLs, snippets, readable page text when fetching succeeds, available page publication/update dates, and the server retrieval timestamp. deep_search reads more sources. If a page is blocked or cannot be extracted, its search result remains a snippet lead; the tool does not present that snippet as verified page text. The model should cite only the URLs returned by the tool.

For a factual question, the model should call smart_search before answering. Time-sensitive queries include date context, selected pages are read, and the result includes a retrieval timestamp. The timestamp means “fetched at”; it does not mean that every result was published or updated then. When available, page publication and update dates are labeled separately.

For release dates, schedules, and other events that may have changed, use evidence published after the scheduled date to decide whether the event actually happened. If only older announcements are available, report what was scheduled at the time and say the outcome could not be verified. Prefer official announcements for plans and independent current reporting for what happened. Cite only URLs returned by the tools; never turn a search snippet or an old prediction into a confirmed fact.

Search uses the ddgs metasearch library with automatic backend selection by default. If it fails or yields no results, DuckDuckGo HTML is tried. When a requested date filter cannot be honored by that HTML fallback, the result is explicitly marked freshness-unverified. Set MCP_SEARCH_REGION (default in-en) to change the search region, for example us-en or uk-en. MCP_SEARCH_BACKENDS can override the DDGS backend, for example bing,google; backend availability depends on network and provider behavior.

Search requests go to the configured DDGS search backend and may fall back to DuckDuckGo HTML. Source reading contacts the result pages; specialized tools also contact Reddit, YouTube, GitHub, and Open-Meteo. Queries and URLs are sent to those external services as needed. Direct URL fetches bypass environment-configured HTTP proxies so the checked DNS address is the address used for the connection; networks that require a proxy may not be able to fetch pages. The server does not intentionally log query strings or fetched page contents; diagnostic logs go to stderr and identify hosts or error types.

No scraper can guarantee that a page is accessible or that a search engine ranks the best sources first. Some pages block automated requests, hide content behind scripts/paywalls, or omit publication dates. If source pages cannot be fetched, the result says to treat snippets as leads rather than verified evidence.

The server cannot edit LM Studio's original prompt before generation. It supplies date and source context only after the model calls an MCP tool, so the client's system instructions must direct the model to call search before answering current factual questions. Search ranking can be stale, and a retrieval timestamp does not prove result freshness. For scheduled events whose dates have passed, the server asks for post-event evidence and labels status unverified when that evidence is absent; it cannot guarantee that hallucinations are impossible.

System prompt

For LM Studio, copy the instructions from LM_STUDIO_PROMPT.md. It directs the model to search before factual answers, distinguish event dates from fetch dates, use post-event evidence for release status, and disclose uncertainty instead of guessing.

Development

The MCP server runs over stdio. The command configured above is equivalent to:

uv run mcp-internet

Source code is under src/mcp_internet/. Tool registration is in src/mcp_internet/server.py; retrieval and extraction are under tools/ and utils/.

Run the offline test suite and configured lint checks with:

uv run --extra dev pytest
uv run --extra dev ruff check src tests scripts

The normal test suite uses mocked responses and does not require network access or credentials.

Common startup failures: if LM Studio reports that uv cannot be found, install uv or set command to its absolute executable path. If the server exits immediately, confirm Python 3.10+, run uv sync in the checkout, and verify both --directory and cwd point to the folder containing pyproject.toml. MCP stdio must remain attached to LM Studio; do not launch the command in a terminal and redirect its output into the MCP connection.

License

MIT. See LICENSE.

Project health

• Changelog
• Contributing
• Security policy