Local (LM Studio)

Run open models locally through LM Studio's built-in server. Blueprint Stack connects to it directly — no API key, no Node.js, no extra adapter — and everything runs on your own machine. Blueprint Stack drives the Unreal Editor by calling tools, so load a model that supports tool / function calling.

Setup

1. Install LM Studio

Download and install LM Studio from lmstudio.ai/download (Windows, macOS, and Linux).

2. Download a model

In LM Studio, search for and download a model that supports tool / function calling — for example Llama 3.1, Qwen2.5, or a similar instruction-tuned model. Smaller models (1B–3B) often struggle with tool calls and may loop or ignore tools; a 7B+ instruction-tuned model is a safer starting point.

3. Start the local server

Open the Developer tab in LM Studio, load your model into the server, and click Start Server. The server listens on http://127.0.0.1:1234 by default. To check it's reachable, open http://127.0.0.1:1234/v1/models in a browser — you should see a JSON list of the loaded model(s).

4. Configure it in Blueprint Stack

Open the Blueprint Stack panel in Unreal, go to Settings > LM Studio, and set:

• -Host — leave blank to use http://127.0.0.1:1234, or enter your address if you changed LM Studio's server port or run it on another machine.
• -Default model — optional. Leave it blank to let LM Studio use whichever model is loaded into the server, or enter a model id to request a specific one. You can also switch the model from the dropdown at any time, including mid-conversation.

5. Use it

Pick Local (LM Studio) from the agent picker, choose a model from the dropdown (the list comes from LM Studio's server), and start chatting. The agent reads your project, calls editor tools, and reports back like the cloud agents — running entirely on your machine.

Models

Whatever model you have loaded in LM Studio's server shows up in the model dropdown. For driving the editor, prefer instruction-tuned models with solid tool-calling support, such as:

• -
  Llama 3.1 (8B / 70B)Good general-purpose default with reliable tool calls.
• -
  Qwen2.5 (7B / 14B / 32B)Strong instruction following and tool use across sizes.
• -
  Other 7B+ instruction-tuned modelsMost modern tool-calling models work; quality scales with size and your hardware.

Agent quality depends on the model and the hardware you run it on. If a model keeps misusing tools or going in circles, try a larger or different one, or raise the model's context length in LM Studio.

Requirements

• -LM Studio installed, with its local server started
• -A model downloaded and loaded into the server (ideally one that supports tool / function calling)
• -Enough RAM / VRAM for the model you choose
• -No API key, no Node.js, no internet connection (once the model is downloaded)

If it can't connect

If your first prompt returns “Couldn't reach LM Studio at http://127.0.0.1:1234”, the server isn't running or Blueprint Stack is pointed at the wrong address:

• -In LM Studio, open the Developer tab, load a model, and click Start Server, then resend the prompt.
• -Confirm http://127.0.0.1:1234/v1/models loads in a browser. If you changed the port, update the Host field in Blueprint Stack settings.
• -Make sure a model is actually loaded into the server (not just downloaded) — the server reports an empty model list otherwise.