# Choosing model routes

Four routes, one picker, mixed freely; every surface treats them identically.

## The four
1. CLI harnesses (claude-cli / codex-cli): your subscription + sign-in, no downloads, no GPU; Kaptain's toolbelt reaches them automatically.
2. Ollama: easy local; auto-started when installed; pull by name.
3. Built-in llama.cpp: direct GGUF control; Kaptain can build the runtime for your GPU.
4. Cloud APIs: OpenAI, Gemini, DeepSeek, Groq, Mistral, Together, OpenRouter, custom OpenAI-compatible — you type the model id your account has.
How to add each: https://chynj.ca/kaptain/docs/how-to/models/

## Spend strength where it matters
Strong model for ambiguous reasoning + final judgment; cheaper routes for mechanical, reviewable work. Delegation automates this (brain routes suitable work to a local worker). Fifth option: another machine's models over the mesh (/knowledge/model-lending/).

## Per-model knobs (gear)
Thinking display, streaming, context window (real default/max shown), max output, tool approval, tool-call style, reasoning effort (minimal..high, discovered from your installed CLI). Small local models need a deliberate num_ctx: 8k with tools is one failed round from full.

## No lock-in
Everything moves later: Settings, per kapsule, per chat; default provider is one field.