Ollama MCP integration

It bypasses the model's prompt templating so the prompt is sent exactly as written, which is useful for debugging or custom formatting. The trade-off is that raw mode does not return a context, so a follow-up cannot continue from it - multi-turn exchanges belong in Chat with Ollama model instead.

So existing code does not have to be rewritten. The OpenAI-compatible chat, text completion, and model list actions accept and return the shapes that API uses, while the native actions expose Ollama's own behaviour such as raw mode. Both hit the same locally installed models.

No. The connected actions run inference and read the inventory - List Models, Show Model Information, and Get Ollama Version - with no action to download, create, or delete a model. Adding a model stays an operation someone performs on the host itself.

No. Ayro still runs the conversation and hands off to a human agent with full history when it cannot resolve something; the Ollama actions are called as tools within that flow. A self-hosted model handles the specific step it is asked to handle rather than replacing the agent.