# LM Studio Use Producer Pal completely offline without an Internet connection. ::: warning Experimental This requires a relatively new machine with decent specs (Apple Silicon with lots of RAM or PCs with Nvidia 4080+ graphics cards). It requires more technical know-how to setup and debug. The online options work significantly better and faster at the time of writing. However, completely offline and private usage is compelling. ::: ## Requirements - [Ableton Live 12.3+](https://www.ableton.com/live/) with [Max for Live](https://www.ableton.com/live/max-for-live/) - [LM Studio](https://lmstudio.ai/) ## Installation Steps ### 1. Install the Max for Live Device Download [Producer_Pal.amxd](https://github.com/adamjmurray/producer-pal/releases/latest/download/Producer_Pal.amxd), the Producer Pal Max for Live device, and add it to a MIDI track in Ableton Live: Producer Pal device running in Ableton Live _It should display "Producer Pal Running" or something isn't working._ ### 2. Enable Small Model Mode (Optional but Recommended) Small model mode provides a smaller, simpler interface optimized for small/local language models. If you're using Option A below, the `-s` flag in the configuration automatically enables small model mode on the device when connected. Alternatively, you can enable it manually in Producer Pal's "Setup" tab: Small model mode setting ### 3. Install a Compatible Model Install a compatible model in LM Studio, such as a recent version of: - Qwen - Gemma - Mistral Magistral or Devstral - GLM Flash ### 4. Configure LM Studio Setup Producer Pal in LM Studio Settings → Program → Integrations → edit mcp.json: **Option A: With npx (recommended)**: ```json { "mcpServers": { "producer-pal": { "command": "npx", "args": ["-y", "producer-pal", "-s"] } } } ``` The `-s` flag enables small model mode. It automatically configures the device when connected, so you don't need to enable it in both places. The producer-pal package is a proxy that responds to requests even when Ableton Live or the Producer Pal device are not running, to let you know there's a problem. **Option B: Direct HTTP**: ```json { "mcpServers": { "producer-pal": { "url": "http://localhost:3350/mcp" } } } ``` ### 5. Verify Tools Confirm the Producer Pal tools are listed under Settings → Program: Producer Pal tools listed in LM Studio _See below for tips on using a subset of tools._ ### 6. Start Using Producer Pal 1. Start a conversation with Producer Pal MCP active and say "connect to ableton" 2. If you didn't enable "Allow all" under Settings → Program, then allow Producer Pal tools in the conversation: ![LM Studio tool permission prompt](/img/lm-studio-permissions.png) ![LM Studio successfully connected to Producer Pal](/img/lm-studio-success.png) ## Local Model Tips - A **context length of about 8k (8000)** is needed to start a conversation with Producer Pal and send a few messages. - A **context length of 16k (16000) or higher is recommended** for achieving useful results, but higher context lengths can make the model run significantly slower, especially as the conversation grows. Experiment to find the right balance. - To help the model make good tool choices and get more out of the limited context length, disable some of the Producer Pal tools. To focus on MIDI clip generation, a good minimal toolset for experimentation is: - `ppal-connect` - `ppal-read-live-set` - `ppal-read-track` - `ppal-create-clip` - `ppal-delete` - `ppal-playback` Try disabling all the other tools and add back when you miss features. - Experiment with empty or extremely simple Live projects - Only work with very simple material (e.g. basic MIDI patterns in clips 4 bars or shorter) - Note that small models are guided to delete and start over rather make edits (other than simple additions) - If the AI struggles and makes mistakes, don't hesitate to delete recent messages from the chat, edit your last message, and try again. Don't waste tokens correcting the LLM. Back up and avoid the issue or try something else. - Shut down every other app you don't need to be running - Consider running Ableton Live on a different machine on the local network - When using a model with the GGUF engine, try enabling the advanced/experimental settings for Flash Attention and setting the K/V caches' quantization to Q8 or Q4. - Research how to optimize for your specific machine / GPU hardware ## Advanced: Customizing Skills Local models vary widely in capabilities. **First, try enabling "Small Model Mode" in Producer Pal's Setup tab** (see installation step 2 above) - this uses a simplified instruction set designed for smaller models. If Small Model Mode isn't working well for you, or you want to experiment further, you can customize Producer Pal's behavior for your specific model: 1. Download this repository and follow the dev setup and build instructions in [DEVELOPERS.md](https://github.com/adamjmurray/producer-pal/blob/main/DEVELOPERS.md) 2. Edit `src/skills/basic.js` - the skills file used by Small Model Mode (or edit `src/skills/standard.js` to adjust regular mode) 3. Experiment with instruction wording, remove features your model struggles with, or adjust the guidance 4. Rebuild with `npm run build` 5. Use the development version of `Producer_pal.amxd` in Ableton Live 6. Reload your Producer Pal MCP server in LM Studio and start a new conversation **Share your findings:** If you discover configurations that work well for specific models, please share in [GitHub Discussions](https://github.com/adamjmurray/producer-pal/discussions). The community benefits from learning what works with different local models. ## Alternative: Built-in Chat UI You can also use LM Studio as a backend for [Producer Pal's built-in chat UI](./chat-ui-other-providers#lm-studio) instead of LM Studio's native interface. ## Troubleshooting If it doesn't work, see the [Troubleshooting Guide](/support/troubleshooting).