A local AI setup you can reproduce.

Start with the task.

A coding assistant, a document workflow, and an image workflow need different tools. Write down what you actually want to run before choosing a model. Save one small input and the result you expect; it gives you a useful check when you change something later.

For a local language-model server, Ollama and llama.cpp are two options with documented APIs. ComfyUI uses a workflow graph for media tasks. RMA keeps your notes and links together; the runtime still runs on your own machine.

Record the exact configuration.

A model name alone is not a setup. Include the model revision or file, quantization, runtime version, launch command, operating system, GPU and VRAM. If a value is unknown, leave it marked unknown instead of borrowing a benchmark from another machine.

Context length, output length, and concurrent requests can change memory use and speed. Keep the conditions alongside any measurements, including whether a run started cold or reused a loaded model.

# My local AI setup
Task:
OS:
GPU:
VRAM:
Model:
Model revision / file:
Quantization:
Runtime and version:
Context length:
Max output:
Concurrent requests:
Launch command:
Test input:
Observed result:
Known limits:
Source URL:

Keep the working version.

Paste the template into Setups in your workspace. Fill in the fields you know, save it privately, and download the text when you need it elsewhere. The importer reads explicit fields; it does not invent hardware specs or performance numbers.

When you try a new model or command, keep a separate setup until you have checked it. Put longer plans in Projects and add the local application URL under My tools. Tool links open applications you already run; they do not install or start a runtime.

Keep credentials out of the record.

Remove API keys, bearer tokens, private connection files, and passwords before saving or sharing a setup. Replace secrets with environment-variable names. The workspace rejects some recognizable secret patterns, but that is not a complete secret scanner.

A successful save confirms that your notes were stored. It does not confirm that the command ran or that a model will fit your hardware. Verify those in the actual local tool.