String Input
input
Team sync — Thursday - Launch date moved to March 14 (was March 7) to finish the accessibility fixes. - Priya owns the onboarding rewrite;…
NodeTool Cloud is in alpha. Direct films and run workflows in the browser, on your own keys.NodeTool Cloud is in alpha. Try it →
Run a workflow on an open model on your own computer. Below is one run on Ollama, the memory each model needs, and what still works when you disconnect.
One local run
The Private Assistant template answers from the text you paste in and nothing else. It is built for a local Ollama model, so the answer comes from the computer you are sitting at.
Install Ollama and pull a chat model, or download one from the Model Manager in Studio. This is the only step that needs the network.
Open the Private Assistant template and choose the Ollama model on its Agent node. The graph has no API key field to fill in.
The Prompt node joins the notes, the question, and the tone. The Agent node sends them to Ollama on 127.0.0.1:11434, and the answer comes back to the Output node.
The inputs as they ship with the template, shortened. Both answers are in the notes: the March 14 launch and the untested billing migration. Paste your own notes to replace them.
What it needs
Memory means graphics memory, or unified memory on an Apple Silicon Mac. Ollama also runs on the CPU, more slowly. The Model Manager in Studio compares these figures with your machine before you download.
| Model on Ollama | Download | Memory | Note |
|---|---|---|---|
| Qwen3.5 2B | 2.7 GB | 4 GB | Runs without a GPU |
| Qwen3.5 9B | 6.6 GB | 8 GB | Chat and tool use |
| GPT-OSS 20B | 14 GB | 16 GB | NodeTool's default local chat model |
| Qwen3.6 27B | 17 GB | 20 GB | The largest of the four |
Approximate figures from the Model Manager. GPU requirements for image, speech, and video models.
What stays on the machine
With the network off
Yes. Each node has its own model. Keep the private step on a local model and send only the step that needs a hosted model to that provider. Only that node needs the network.
Yes. Image and speech models such as Stable Diffusion XL, FLUX.1 schnell, Whisper, and Kokoro run through the HuggingFace pack on an NVIDIA card, an AMD card on Linux, or an Apple Silicon Mac. Local video generation needs an NVIDIA or AMD card. The Model Manager lists the models that fit your machine.
No. Cloud runs on our servers and uses hosted providers. Local models need Studio, the desktop app.
Download the open-source studio for macOS, Windows, and Linux. Open the template and direct it yourself, on your own keys.