How-to guide

Run a model on your own computer

Halo can use a model running on your own Mac or PC through Ollama, LM Studio or any server that speaks the OpenAI format. Then nothing you ask leaves hardware you own.

Set up the server

  1. Install a model server

    Install Ollama or LM Studio and download a model that supports tool calling. Halo’s agent relies on tools, so pick a recent, capable model.

  2. Give it an HTTPS address

    iPhone apps connect most reliably to secure (HTTPS) addresses. The easiest way to get one for your computer is Tailscale: install it on the computer and your iPhone, then share the model server on your private network.

    # Ollama (port 11434)
    tailscale serve --bg 11434
    
    # LM Studio (port 1234)
    tailscale serve --bg 1234

    Tailscale prints an address like https://my-mac.tailnet-name.ts.net. It works from anywhere your iPhone has Tailscale on, not just at home.

Add it to Halo

  1. Add a custom provider

    Go to Settings›AI Model, tap +, and choose OpenAI Compatible API under Custom Provider.

  2. Fill it in

    FieldValue
    ProviderOllama, LM Studio, or any name you like
    API KeyLeave empty
    Custom API Basehttps://my-mac.tailnet-name.ts.net
    Auto Append /v1On
    API FormatChat Completions
  3. Load models and save

    Tap Fetch Models, set the context window to match how you run the model, turn on Supports Thinking for reasoning models, pick your Primary model, and tap Add Provider.

Mix and match

Local models are great as the Secondary model for summaries and quick lookups, while a frontier model handles chat and memory as Primary. You can also choose a model per task.

Your computer has to be on

Background jobs like memory, insights and scheduled tasks fail while the computer is asleep or out of reach. Set a cloud model as Fallback so Halo keeps working.

Troubleshooting

  • Can’t connect. Open your base URL plus /v1/models in Safari on your iPhone. If it doesn’t load, check that Tailscale is connected on both devices and the model server is running.
  • The model ignores tools or loops. Use a larger model with tool calling support, or raise its context window.