# Skill: Set up LM Studio (or an equivalent GUI) to run local LLMs without the terminal

> Apply this skill when a user wants to run a local AI model but prefers a graphical interface over command-line tools, or when they need a local OpenAI-compatible server without writing any commands.

Source: sodigi-learn -- local-agents/lm-studio-i-gui -- https://sodigi.io/learn/local-agents/lm-studio-i-gui

## When to use
- User is intimidated by the terminal and wants a point-and-click workflow for local AI.
- User needs to browse, compare, and download models with visible memory-fit warnings before committing to a download.
- User wants a built-in offline chat that works like ChatGPT but runs on their own hardware.
- User wants to expose a local OpenAI-compatible API endpoint to connect editors, agents, or scripts.

## Core rules
- LM Studio is a GUI wrapper (remote control) for a local inference engine -- it is not the model itself.
- After installing LM Studio, the user must still download at least one model via the Discover tab before any chat is available.
- The Discover tab searches Hugging Face and shows memory requirements per variant before downloading.
- Always pick a quantized variant (start with the 4-bit option) that fits in available memory with headroom.
- The local server toggle (Start server) exposes an OpenAI-compatible API on localhost; this is the bridge to agents, editors, and scripts.
- Swapping from the cloud to the local server requires changing only one field in client code or settings: the base URL.
- Alternatives: Jan (fully open source, same concept), Ollama (terminal-first, for script automation and minimal footprint).
- Models downloaded from inside the program come through the trusted Hugging Face catalog; do not import random files from external sources.

## Procedure
1. Download LM Studio from its official website and install it like a normal application (Windows, macOS Apple Silicon, Linux).
2. Open the program and go to the Discover tab (magnifying glass icon).
3. Search for a beginner-friendly instruct model (e.g., a small Llama or Qwen instruct variant).
4. Read the memory hint next to each variant; choose the quantization level that fits the user's VRAM with at least 2 GB to spare -- typically the 4-bit (Q4) variant.
5. Click Download and wait for it to finish.
6. Switch to the Chat tab, select the downloaded model at the top, and send a test message to confirm it responds.
7. To expose the API: open the Local Server tab and flip the Start server toggle. Note the port number shown.
8. In any client tool, set the base URL to the address LM Studio displays (e.g., `http://localhost:1234/v1`) and use any dummy string as the API key.
9. Verify connectivity by checking the server's request log in LM Studio while sending a test request.

## Ready-to-use prompt
```
Help me get a local AI model running in LM Studio (no terminal needed).

My hardware:
- OS: [Windows / macOS with Apple Silicon / Linux]
- GPU VRAM (or total RAM if no GPU): [X GB]

Please:
1. Confirm I should install LM Studio from its official site (not a third-party source).
2. Recommend a specific model family and size that fits my memory in Q4 quantization.
3. Walk me through the Discover tab -- search, memory hint, and download steps.
4. Explain how to open the chat and confirm the model is working.
5. Show me how to start the local server and what base URL + dummy key to use in another app.
6. List the common mistakes to avoid on first run.
```

## Pitfalls
- Installing LM Studio and expecting a working chat immediately -- you must download a model first; the GUI alone is an empty shell.
- Downloading the largest model variant "to get the smartest result" -- it will not fit in memory or will be unusably slow.
- Ignoring quantization variants -- the same model comes in multiple sizes; always check the memory hint and pick Q4 for a first run.
- Activating the local server and then forgetting to keep it running -- a toggled-off server looks like a broken connection from the client side.
- Downloading models from sources outside the Discover tab -- use the built-in catalog to stay on trusted, verified files.
- Assuming LM Studio is required -- if the user prefers the terminal or needs lightweight scripting, Ollama is the better fit.