Claude Model Router Hook

Effort-first model routing for Claude Code. Heuristics-first, opt-in autoswitch.

Install

installer="$(curl -fsSL https://raw.githubusercontent.com/tzachbon/claude-model-router-hook/361031c9381fd6d2b5c47b311b7037015b3c2f6d/install.sh)" && bash -c "$installer"

The wizard asks whether to install globally and star the GitHub repository. Both prompts default to yes. Installation requires Claude Code and Python 3; starring requires an authenticated gh command. Restart Claude Code after installation.

Claude Model Router classifies each prompt by task and effort, then recommends the matching Claude model. It warns by default. Autoswitch is opt-in and only affects new sessions.

What it does

• Routes prompts with local heuristics and an optional Claude Code fallback.
• Uses Haiku only for clearly mechanical work; defaults agentic coding to Opus with graduated effort.
• Routes Agent and Task sub-agents when they spawn. Nested delegation is floored at the implementation target, and an explicit model always wins.
• Records the requested and resolved model for completed sub-agents, so host-side fallback or allowlist substitutions are visible.
• Dampens borderline classifications instead of switching tiers on weak evidence.

How it works

Hook	Job
UserPromptSubmit	Classifies the prompt and warns or writes the next-session model.
PreToolUse	Routes Agent and Task spawns.
PostToolUse	Logs the model actually used by completed Agent and Task work.
SessionStart	Adds the shared task-class rules to each session.

Errors fail open and leave the prompt unchanged. Prefix a prompt with ~ or < to skip routing.

Model Tier Rules

These rules apply to YOU and to every sub-agent you spawn.

Task classes and default targets

Class	Target model	Effort	When to use
mechanical	haiku	none	Git ops, renames, formatting, lint, file moves, version bumps, quick lookups, short imperative tasks.
implementation	opus	medium	Writing or editing code, building features, creating components or APIs, writing tests, standard feature work.
debugging	opus	high	Diagnosing failures, flaky tests, races, regressions, stack traces, bisecting, reproducing bugs.
architecture	opus	xhigh	Architecture decisions, tradeoff analysis, redesigns, deep multi-file analysis, sustained reasoning over large context.
extreme	opus	max	Multi-system migrations, codebase-wide rewrites, long-horizon plans, RFCs and design docs, platform-scale work.
abstain	(no routing)	-	Prompt does not clearly match any class; current model and effort pass through unmodified.

Sub-agent model selection (MANDATORY)

When calling the Agent tool, set the model parameter to match the task class above. Do not default every sub-agent to the highest effort. Match the model and effort to the work: mechanical work goes to haiku, implementation to opus, debugging to opus, deep analysis to opus, and platform-scale work to opus.

Configure

Use ~/.claude/model-router.json globally or .claude/model-router.json in one project. Project settings win.

{
  "version": 2,
  "apply_mode": "warn",
  "allow_fable_autoswitch": false,
  "subagent_enforcement": "on",
  "classifier": { "cli_fallback": true },
  "thresholds": { "effort_warn_distance": 2 }
}