Local models

nod can use a model that runs on your own computer, or on any server with an OpenAI-compatible API. Model requests then go only to the address you enter, and a local model costs nothing per request.

Local models work in the desktop app and source installs.

Your model also decides when nod asks you

nod's trust check runs on your selected model. It decides whether your words cover each app action, website step and sign-in, or whether nod asks you first. Smaller models make that call less reliably: they can ask more often, or run something you didn't ask for. Use the largest model your computer runs well.

Before you start

You need two things from the model server:

• A model that supports tool calls. nod does all its work through tools, so a model without them can only chat.
• At least 64,000 tokens of context. nod's instructions and tools take about 20,000 tokens before your first message. Most servers start with 4,096, and some cut nod's instructions off without an error.

Set up your server

Ollama

1. Install Ollama and pull a model that supports tools. On ollama.com, filter models by Tools.

   Terminal

   ollama pull gpt-oss:20b

2. Start Ollama with a larger context. In the Ollama app, set Context length in Settings instead.

   Terminal

   OLLAMA_CONTEXT_LENGTH=65536 ollama serve

3. Use the address http://localhost:11434/v1 in nod.

nod reads each Ollama model's details. It lists only models that support tools, sends images to vision models, offers the model's thinking levels and uses its context length.

LM Studio

1. In LM Studio, download a model that supports tools.
2. Load it with Context Length set to 65,536 or more.
3. Start the server from the Developer tab, or run lms server start.
4. Use the address http://localhost:1234/v1 in nod.

With LM Studio 0.4 or later, nod reads each model's details. It skips embedding models and models not trained for tool use, sends images to vision models and uses the context length the model is loaded with, so the model list shows when that's too small.

vLLM

1. Start vLLM with tool calls on. Choose the tool call parser that matches your model.

   Terminal

   vllm serve <model> --max-model-len 65536 --enable-auto-tool-choice --tool-call-parser <parser>

2. Use the address http://localhost:8000/v1 in nod. If you started vLLM with --api-key, add that key too.

llama.cpp

1. Start llama.cpp's server. --jinja turns on tool calls.

   Terminal

   llama-server -m <model>.gguf -c 65536 --jinja

2. Use the address http://localhost:8080/v1 in nod.

Connect nod

1. Open the account menu, select Models, then select Add server next to OpenAI-compatible. During onboarding, pick the OpenAI-compatible card instead.
2. Check the Server address. nod fills in Ollama's address, http://localhost:11434/v1. If you leave out /v1, nod tries it for you.
3. Add an API key only if the server needs one. During onboarding, select Add an API key first.
4. Select Save connection, or Connect during onboarding.

nod checks the address by listing the server's models. The connection shows the server's host, such as localhost:11434. If you haven't chosen a model yet, nod selects the server's first one. To switch, open Settings → Models and choose a model.

After you pull or load another model, select the refresh button next to the connection. To change a server's key, add the same address again: nod updates that connection instead of adding a second one.

Set the server in .env

In a source install, you can set the server in .env instead of in the app. It then shows Server settings in Models, and you remove it from .env.

.env

LLM_PROVIDER=compatible
OPENAI_COMPATIBLE_BASE_URL=http://localhost:11434/v1
OPENAI_COMPATIBLE_API_KEY=
LLM_MODEL=gpt-oss:20b

Leave OPENAI_COMPATIBLE_API_KEY empty if the server doesn't need a key, and LLM_MODEL empty to use the server's first model. Restart nod after you edit .env. If the server isn't running when nod starts, start it and select the refresh button next to the connection. See Configuration.

Other OpenAI-compatible services

The same connection works with hosted services that offer an OpenAI-compatible API. Enter the service's address and your API key.

Service	Address
Groq	https://api.groq.com/openai/v1
Together AI	https://api.together.xyz/v1
DeepSeek	https://api.deepseek.com/v1

nod lists every model such a service returns, including ones that can't chat. Pick a chat model that supports tools.

What changes with these connections

Area	With a local or OpenAI-compatible model
Spending	nod doesn't know the server's prices, so it records each request at $0. That usage doesn't count toward the monthly limit or task check-ins. A paid service still bills you.
Images	Ollama and LM Studio models that support vision get screenshots and images. Models on other servers get text only.
Thinking levels	Only Ollama models offer them. Other servers use the model's own default.
Decision checks	Your selected model reads your replies to confirmations and checks emails that give account access. See Decision checks.
Background work	Memory upkeep and other background work use your selected model.
Where requests go	Model requests go only to the address you enter. Web search, websites, your apps and chat channels work as before. See Your data and privacy.
Where your key goes	An API key stays in nod's encrypted key store and is added to each request by the nod server, like any other model key. See Where your keys go.

Troubleshooting

Problem	Fix
nod couldn’t list this server’s models	Check that the server is running, and that the address and API key are right. curl http://localhost:11434/v1/models should list your models.
This server has no models nod can use	nod needs a model with tool support, and from Ollama it lists only those. Pull or load one, then try again.
nod ignores your request, forgets its tools or repeats itself	The context is too small. Raise it to 65,536 or more and reload the model.
A new model doesn't appear	Select the refresh button next to the connection.
Can’t load its models next to a server set in .env	Start the server, then select the refresh button next to the connection.
Remove it and add the server again	nod can't read the connection's saved settings. Remove the connection, then add the server again in Models.
nod asks before almost everything, or acts without asking	The model is too small for the trust check. Switch to a larger model.

Related resources

• Models
• Approvals and the trust check
• Spending
• Your data and privacy