Use a local model with AI Coach

PulseLoop can run AI Coach conversations through an OpenAI-compatible server you control. The server can run on a Mac, PC, NAS, or another computer on your network, and most setups do not need an API key.

Use the server's network address

If the model runs on another computer, do not enter localhost or 127.0.0.1 in PulseLoop. On your iPhone, those addresses mean the iPhone itself. Use the computer's LAN address instead, such as http://192.168.1.50:11434.

Quick start

1. Start one of the supported servers, load a chat model, and make the server reachable from your local network. Keep the iPhone and server on the same network for the first setup.
2. In PulseLoop, open Settings → AI Coach and turn on Enable AI Coach.
3. Set Provider to Local / self-hosted.
4. Enter the server's base address, for example http://192.168.1.50:11434. You do not need to add /v1. Add an API key only if your server requires one.
5. Tap Detect server & configure. PulseLoop checks the connection, lists available models, and tests tool calling and structured output. Choose a model if the server offers more than one.
6. Open the Coach and try a question such as “How was my sleep last night?” Allow Local Network access if iOS asks.

Model size and context

Tool-capable instruct models work best. A context window of at least 8K is a good starting point; very small contexts can truncate PulseLoop's health context and tool definitions.

Supported servers

PulseLoop uses the OpenAI-compatible Chat Completions API. These engines are supported directly:

Engine	Typical address from the iPhone	API key	Setup notes
Ollama	http://<server-ip>:11434	Usually none	Tool use depends on the model. Ollama listens only on the server itself by default; enable LAN access with OLLAMA_HOST=0.0.0.0:11434.
LM Studio	http://<server-ip>:1234	Optional	Start the local server and enable Serve on Local Network. Authentication is off by default.
llama.cpp (llama-server)	http://<server-ip>:8080	Optional	Use --host 0.0.0.0 for LAN access. --jinja is recommended for tool-capable models.
vLLM	http://<server-ip>:8000	Optional	Automatic tool calling needs --enable-auto-tool-choice and the parser required by the model.
SGLang	http://<server-ip>:30000	Optional	Tool and structured-output support depends on the model and server configuration.
Other OpenAI-compatible servers	Your server's base address	Varies	The server must provide GET /v1/models and POST /v1/chat/completions.

The port can be different if you changed the server configuration. PulseLoop also accepts a full /v1 or /v1/chat/completions URL and normalizes it to the base address.

Recommended settings

• Detect server & configure: Run this first and again after changing the server, model, or launch options.
• Tool calling: Keep this on for questions that need detailed PulseLoop data. It also powers goals, workout logging, meal logging, and other actions when AI actions is enabled. Turn it off if the server or model does not support tools reliably.
• Response format: The detected setting is usually best. Off is the broadest-compatibility fallback if a server returns schema or JSON errors.
• Max tokens: Leave this blank unless the server needs an explicit limit.
• Timeout: The 180-second default suits most local models. Increase it for a large model running on a CPU; use a smaller model if every response times out.
• API key: Leave it blank unless authentication is enabled on your server. Saved keys are stored in the iOS Keychain.

Local servers can use PulseLoop's health-data and action tools. Image input also works when the selected model supports images. Provider-hosted web search is not available in local mode.

Troubleshooting

Problem	What to check
Server not reachable	Use the server computer's LAN IP, not localhost. Confirm both devices are on the same network and allow PulseLoop's iOS Local Network permission.
Connection refused	Make sure the server listens on 0.0.0.0 or its LAN interface, and allow its port through the computer's firewall.
No models found	Load a model, then open <base-address>/v1/models from another device on the same network. It should return JSON.
Coach answers but cannot read or change data	Enable Tool calling, use a tool-capable model, and check the engine-specific launch options in the table. Enable AI actions for writes.
JSON or schema errors	Set Response format to Off, then run detection again.
Empty, truncated, or timed-out answers	Use at least an 8K context window, leave Max tokens blank, increase the timeout, or choose a smaller model.
401 / unauthorized	Save the same API key or token configured on the server.

Privacy and network access

The local provider sends your question and the health context needed to answer it to the server you configure. Use a server and network you trust.

PulseLoop accepts plain http:// addresses for local/private destinations and requires https:// for public hosts. Redirects are not followed, and an optional API key is stored in the iOS Keychain. Do not expose an unauthenticated local model server directly to the public internet.