Skip to main content

​
Ollama

Ollama runs AI models locally — no API key, no cloud, no data leaving your network. WednesdayAI auto-discovers a running Ollama instance.

​
Prerequisites

Install Ollama from ollama.com, then pull a model:

​
Connecting

WednesdayAI detects a running Ollama instance automatically. To activate auto-discovery, set OLLAMA_API_KEY to any non-empty value — the value is not validated against any service, it just signals OpenClaw to enable the Ollama provider:
Once set, WednesdayAI probes http://127.0.0.1:11434 by default. To point at a non-default host, use models.providers (see Remote Ollama) — OLLAMA_HOST is an Ollama-native variable for the Ollama process itself and has no effect on OpenClaw activation.

​
Selecting a model

The ollama/ prefix routes the request to the Ollama provider: Run openclaw models status to see which Ollama models WednesdayAI can reach.

​
Remote Ollama

Do not append /v1 to the baseUrl. Using http://host:11434/v1 routes through the OpenAI-compatibility shim and breaks tool calling. Always use the native Ollama URL: http://host:11434.
If Ollama runs on another machine (a GPU box, a home server), set the base URL via models.providers:
By default Ollama only listens on 127.0.0.1. To accept remote connections, set OLLAMA_HOST=0.0.0.0 on the Ollama machine. For secure remote access, put Ollama behind Tailscale and use its MagicDNS name rather than exposing port 11434.

​
Troubleshooting

Ollama is running but no models are available, OR the models you have installed don’t report tools capability. WednesdayAI only auto-discovers models that report tools support in /api/show. To check:
Look for tools in the capabilities array. Pull a model with tool support (e.g., ollama pull llama3.2) or explicitly configure the model in models.providers.ollama.models.
Ollama is not running. Start it with ollama serve (or the system service) and verify with ollama list.
By default Ollama listens only on loopback. Set OLLAMA_HOST=0.0.0.0 on the Ollama machine and restart it. Use Tailscale rather than exposing the port publicly.