NemoClaw Knowledge Wiki

Using LM Studio completely locally for web browsing

3 min read

Using LM Studio completely locally for web browsing

https://www.youtube.com/watch?v=kKNgRCPuObI Here is a Markdown summary of the video tutorial on using LM Studio with the Model Context Protocol (MCP).

Turning LM Studio into a Local AI Command Center with MCP

This video demonstrates how to transform LM Studio from a simple text-based chat interface into a powerful, automated agent capable of browsing the web, checking news, and using maps—all running locally, privately, and for free.

1. What is an MCP?

MCP stands for Model Context Protocol.

  • The Problem: Standard LLMs (Large Language Models) can predict text but cannot take actions or access the outside world.
  • The Solution: MCP is a universal standard that allows AI models to connect to external tools (browsers, databases, GPS, etc.) safely and structurally.
  • How it works: The model describes what it wants to do, the tool executes the code, returns the data, and the model uses that data to determine the next step.
  • Key Benefit: It is not a proprietary plugin system; it is an open protocol compatible with many AI tools.

2. Initial Setup

  1. Download LM Studio: Get the software from lmstudio.ai.
  2. Select a Model: You must use a model capable of “Tool Use”.
    • Recommendation from video: Qwen 2.5-VL-30B (This model supports both vision and tool use).
  3. Configuration: Ensure you maximize the context window length in settings to allow enough space for tool data.

3. How to Install MCPs

LM Studio has an “Integrations” tab (represented by a plug icon) where you edit a file called mcp.json. You add specific configurations here to enable new tools.

Demo A: Web Browsing (Playwright)

To give the AI access to the live internet:

  • Tool Used: Playwright (an open-source tool for automating web browsers).
  • Installation: Added via npx command in the mcp.json file.
  • Capabilities:
    • Navigate to websites.
    • Read headlines and content.
    • Vision Capabilities: The model can be instructed to take a screenshot of a webpage and display it directly in the chat window.

Demo B: RSS Reader

To fetch structured news feeds:

  • Tool Used: An RSS Reader MCP.
  • Workflow: The user asked the AI to go to news.ycombinator.com (Hacker News) via RSS to get the top posts.
  • Agentic Behavior (Chaining): The video demonstrated combining tools:
    1. Used RSS to find the top 20 articles.
    2. User asked for content of specific article (#4).
    3. AI switched to Playwright to visit that specific link and scrape the content.

Demo C: Google Maps

To find real-world location data:

  • Tool Used: Google Maps MCP.
  • Requirement: Requires a Google Maps API Key (users must generate their own).
  • Capability: The AI can perform geolocation, search for places (e.g., “coffee shops near Rittenhouse Square”), and retrieve ratings, addresses, and hours of operation without leaving the local chat.

4. Where to Find More Tools

The video recommends visiting mcp.so.

  • This is a directory of MCP servers and clients.
  • You can find tools for databases (SQLite), productivity (Slack, GitHub), and utilities (Time, Filesystem).
  • Installation: Find the tool you want, copy the JSON configuration provided on the site, and paste it into your LM Studio mcp.json file.

Key Takeaways

  • Local & Private: All orchestration happens on your computer.
  • Agency: By chaining MCPs, you can create complex workflows (e.g., “Find a news story, go to the website, take a screenshot”).
  • Future-Proof: Since MCP is a standard, tools built for other platforms (like Claude Desktop) will generally work with LM Studio.

Graph View