Skip to main content

Model Routing (Task-Based)

Model routing automatically selects the best model for each task type based on the content of the request. This lets you use a powerful model for complex reasoning while routing simpler tasks to a faster, cheaper model.

Brains / Muscles architecture

ClawCentral uses a Brains / Muscles model:

  • Brains — the primary model for complex, reasoning-heavy tasks (planning, coding, analysis). Default: your configured primary model.
  • Muscles — a secondary model for fast, straightforward execution tasks (formatting, lookups, simple transforms). Default: a faster/cheaper model.

The agent selects which model to use based on the detected task type, without any manual intervention.

Task routing categories

Task typeDescriptionDefault model
planningArchitectural decisions, multi-step plansBrains
codingWriting, reviewing, or debugging codeBrains
researchWeb search, information gathering, summarizationBrains
creativeWriting, brainstorming, generation tasksBrains
executionRunning tools, formatting output, simple lookupsMuscles

Routing heuristic

Routing is determined by a smart heuristic that analyzes the message content:

  • Keywords and phrasing patterns signal the task type (e.g., "write a function", "what is", "format this").
  • Tool calls in progress shift routing toward execution.
  • Ambiguous requests default to Brains to avoid degrading quality.

Configure task routing

Task routing is configured in Admin UI → Agents → [agent name] → Task Routing.

Settings:

  • Enable task routing — toggle on/off (off means all tasks use the primary model)
  • Brains model — model for planning, coding, research, creative tasks
  • Muscles model — model for execution tasks
  • Per-task override — pin a specific model to any task category

Example: Use claude-opus-4-5 as Brains and claude-haiku-4-5 as Muscles. Coding tasks go to Opus; tool-execution tasks go to Haiku. Both share the same BYOK API key.

Why use task routing?

  • Cost — execution tasks don't need the most expensive model.
  • Speed — Muscles responses are faster, improving perceived latency for simple turns.
  • Quality — critical reasoning still uses your highest-capability model.

Notes

  • Task routing is per-agent. Different agents can use different routing configs.
  • BYOK API keys apply to both Brains and Muscles; both must be supported by the configured provider.
  • Streaming works the same regardless of which model is selected.
  • If the Muscles model is unavailable, routing falls back to Brains automatically.