Models & Authentication

Nexus-Nancy supports multiple ways to connect to LLMs, from high-performance local servers to cost-effective ChatGPT Plus subscriptions.

1. Local LLM (Gemma 4 / Llama 3)

Ideal for private, high-speed execution.

Start the Llama Server

llama-server -m /path/to/model.gguf --port 8089 --jinja --reasoning-format deepseek --think

Configure Nancy

# .agents/nnancy.yaml
model: my-local-model
base_url: http://localhost:8089/v1
auth_type: api_key
api_key_file: none

2. ChatGPT Plus Subscription ($20/mo)

Nexus-Nancy can bridge directly to your ChatGPT Plus subscription using OpenAI Codex OAuth. This avoids per-token API billing and allows you to use your flat-rate subscription.

Login

Run the following command on your local machine:

nnancy auth login

This will open your browser to OpenAI’s authorization page. Once you log in, Nancy will capture the session tokens and store them securely in .agents/secrets/codex.json.

Configure Nancy

# .agents/nnancy.yaml
auth_type: codex
model: gpt-5.4  # Or latest available via Codex

Nancy will automatically handle token usage and refreshing. Note that this method is subject to your subscription’s message caps (e.g., 80 messages / 3 hours).

3. Standard OpenAI API

Standard usage-based billing.

# .agents/nnancy.yaml
auth_type: api_key
api_key_env: OPENAI_API_KEY
base_url: https://api.openai.com/v1

Setting Up Your API Key

To securely add your API key:

nnancy secrets

This opens .agents/secrets/openai.key in your editor. Write your key to this file and save. The key is read on startup and validated before any API calls.

Alternatively, set the OPENAI_API_KEY environment variable.

Model Switching Strategy

Nancy follows a Minimalist Configuration philosophy. We do not provide commands to switch models (e.g., nnancy use gpt-4). Instead, we encourage users to edit the .agents/nnancy.yaml file directly.

This ensures:

  1. Source of Truth: The config file always reflects the current state.

  2. Explicitness: You always know which model is active and what it costs.

  3. Admin-Free: No complex hidden state to manage across different cluster nodes.