Skip to content

Thinking effort

Milimarty edited this page Oct 19, 2026 · 1 revision

Thinking effort

Each agent, and KeelBot, has a thinking effort next to its model: how hard the model thinks before it acts. More effort is slower and uses more tokens; less is quicker and cheaper. default leaves it to the engine.

  • Where to set it: the Agents page (per agent, per project), KeelBot's model picker (for one chat), keel TUI (/model asks for the effort after the model; /effort high, /effort default), and MCP (keel_agent_model, or keel_agent_create with effort; both ask you first).
  • Where you see it: next to the model everywhere: Agents, Jobs, Live agents, the Flow page, KeelBot's header and each answer ("sonnet · high"), keel TUI's job lines and "✓ Finished … · claude/sonnet · high", and keel_agents. Each agent call keeps the effort it ran with, so the history shows it too.
  • Which models have it:
Engine Efforts How keel passes it
Claude Code (subscription) low, medium, high, xhigh, max (not Haiku) claude --effort
Claude API key low, medium, high, xhigh, max (not Haiku) output_config.effort
Codex CLI and OpenAI key what each model offers (from Codex's model list) -c model_reasoning_effort=… / reasoning_effort
Copilot through OpenCode low, medium, high (newer GPT also xhigh; Gemini none) opencode run --variant
Copilot CLI, GitHub Models key none —
Fake model low, medium, high (only noted, for tests) —

A model without an effort setting shows "no effort setting", and keel ignores an effort it cannot pass on.


From docs/GUIDE.md · All pages

Clone this wiki locally