Ollama — run the AI locally (no cloud, no cost)

The Deputy and Posse can run on a LOCAL model via Ollama instead of the Anthropic cloud — pick the provider in the Deputy's "AI" dropdown (Anthropic (cloud) or Ollama (local)). Local means: no API key, no per-token cost, and nothing leaves your machine. Good for private, air-gapped, or just-experimenting work.

Setup:

1. Install Ollama (ollama.com) and make sure it's running — it serves on http://localhost:11434 by default (you can also point at another machine on your LAN).

2. Pull a TOOL-CAPABLE model: `ollama pull llama3.1` (also good: qwen2.5, mistral). Tool use — the whole point of the Deputy — needs a model that supports function calling; not every local model does.

3. In the Deputy: set AI = Ollama (local), confirm the host, and pick the model (the ↻ button lists installed ones).

How it works: SubnetSlinger talks to Ollama's /api/chat endpoint and translates to/from the same internal shape the Anthropic client uses — so the agent loop, all the read-only tools, and any connected MCP servers work identically. The token line shows in/out and says 'local (free)'. The Posse follows whatever provider the Deputy is set to.

Tradeoffs: local models are private and free but generally weaker at long multi-step tool chains and slower on modest hardware than the frontier cloud models. Use Haiku/Sonnet/Opus for the hardest investigations; use Ollama when privacy, cost, or offline operation matters more. Same read-only safety either way — the AI never changes device state.

Does it get better over time? No — a local model's weights are FIXED, so it does not learn from your usage or build up better predictions the more you use it. It only 'remembers' within the current chat (the context window), and that resets on New chat. To get stronger answers you pull a bigger/newer model or switch to the cloud for that one hard investigation — not by using it more.

Two practical notes: the Ollama app must be RUNNING before you pick Ollama (start it first — the Deputy reaches it at localhost:11434), and the very first reply after a cold start can be slow while the model loads into memory; it speeds up once it's warm. It works — just give that first prompt a moment.

Under the hood this is a provider seam (IChatProvider): ClaudeClient and OllamaClient are two implementations, so a future backend (e.g. OpenAI) would just be one more adapter.