Concept · Models · 4 min read

Choosing model routes

Kaptain is model-agnostic in the useful sense: four genuinely different routes live in one picker, mix freely in one conversation history, and every surface — chat, kapsules, delegation — treats them identically.

The four routes

CLI harnesses (claude-cli, codex-cli, gemini-cli): your existing subscription and sign-in, no downloads, no GPU — and Kaptain's toolbelt reaches them automatically. Ollama: the easy local route; Kaptain auto-starts an installed one and pulls models by name. Built-in llama.cpp: direct GGUF control; Kaptain can build the runtime for your GPU itself. Cloud APIs: OpenAI, Gemini, DeepSeek, Groq, Mistral, Together, OpenRouter, or any OpenAI-compatible endpoint — you type the model id your account actually has. Adding each: How-To ▸ Add models.

Spend strength where it matters

The operating rule: a strong model for ambiguous reasoning and final judgment; cheaper routes where the task is mechanical and the result reviewable. That's what delegation automates — the brain model routes suitable work to a local worker, and the tokens land on hardware you already paid for. A fifth option compounds it: another machine's models over the mesh.

The per-model knobs

Every model carries its own ⚙ settings: live thinking display, streaming, context window (with the runtime's real default and max shown), max output, tool-approval mode, tool-call style, and — where the provider reports levels — reasoning effort from minimal to high, discovered from your installed CLI rather than assumed. Small local models deserve a deliberate context-window choice: an 8k window with tools is one failed round from full.

Nothing is a lock-in

Setup's model step is a front door. Every choice moves later in Settings, per kapsule, per chat — and the default provider is one field, so changing your mind fleet-wide is one edit.

Agent-readable version: index.md