LM Studio setup: Gemma, MCP tools and Claude desktop

This tutorial uses LM Studio to run models locally, starting with installation, hardware detection and model downloads. The presenter explains the difference between storing a model on disk and loading it into memory, then adjusts context length before starting a chat.

Memory headroom matters. The example model is listed at 4.38 GB, while the displayed memory estimate rises to about 6 GB after increasing context capacity. The presenter warns that conversation history and tool calls can add memory demand. He also argues that smaller models may struggle with coding and tool use compared with cloud services; these are his assessments rather than benchmark results.

LM Link connects a MacBook Pro to models on a Mac Studio. The demonstration selects Gemma 4 26B and asks it to describe an uploaded image. For MCP setup, the walkthrough installs Node.js and edits LM Studio's JSON configuration. Brave Search requires an API key and is described as a paid service. Playwright controls a browser without an API key, though the screenshot task includes a failed call and retries.

The final section starts LM Studio's server and copies its local URL into Claude desktop's configuration. The presenter refers viewers elsewhere for the full configuration steps. He describes ordinary local inference as usable offline; the web search and website automation examples involve external services.