claude-model-router-hook

A Claude Code hook system that classifies every prompt by task complexity and switches your active model automatically. Sub-agent model rules are injected into every session so spawned agents also use the right tier.

Features

• Classifies prompts by complexity using keyword and pattern matching (zero API calls)
• Warns on tier mismatch by default; optionally auto-switches settings.json (action: "autoswitch")
• Injects sub-agent model-selection rules into every session via SessionStart
• Prefix any prompt with ~ to bypass classification and keep the current model
• Logs every classification and switch to ~/.claude/hooks/model-router-hook.log

How It Works

Two hook scripts run inside Claude Code:

session-init.sh (SessionStart) injects a systemMessage into every session that enforces these sub-agent rules:

model-router-hook.sh (UserPromptSubmit) classifies the incoming prompt and compares the recommended tier against the current model in settings.json. By default it warns with a recommendation; set "action": "autoswitch" in your config to switch automatically.

Installation

Plugin install (recommended)

claude plugin marketplace add tzachbon/claude-model-router-hook
claude plugin install claude-model-router-hook@claude-model-router-hook

Hooks are registered automatically. Restart Claude Code to activate.

One-liner (manual)

curl -fsSL https://raw.githubusercontent.com/tzachbon/claude-model-router-hook/main/install.sh | bash

Manual

git clone https://github.com/tzachbon/claude-model-router-hook.git
cd claude-model-router-hook
./install.sh

Or copy manually:

mkdir -p ~/.claude/hooks
cp hooks/session-init.sh hooks/model-router-hook.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/session-init.sh ~/.claude/hooks/model-router-hook.sh

Then add hooks to ~/.claude/settings.json (use the full absolute path from echo $HOME):

Under SessionStart:

{
  "type": "command",
  "command": "/home/yourname/.claude/hooks/session-init.sh",
  "timeout": 2
}

Under UserPromptSubmit:

"UserPromptSubmit": [
  {
    "matcher": "",
    "hooks": [
      {
        "type": "command",
        "command": "/home/yourname/.claude/hooks/model-router-hook.sh",
        "timeout": 2
      }
    ]
  }
]

Then restart Claude Code.

Override

Prefix any prompt with ~ to skip classification entirely and keep the current model active.

Configuration

Routing behavior is controlled by ~/.claude/model-router.json (global) or .claude/model-router.json (project-level, takes priority).

See the schema for all options (thresholds, per-tier keywords/patterns, etc.).

Log

Activity is written to ~/.claude/hooks/model-router-hook.log:

[2026-03-07 12:00:00] model=sonnet rec=opus action=WARN->opus prompt="analyze the entire..."
[2026-03-07 12:01:00] model=opus rec=match action=ALLOW prompt="git commit changes"
[2026-03-07 12:02:00] OVERRIDE prompt="~ keep opus for this..."

Setup Prompt

Copy and paste this into any Claude Code session to install the hooks automatically.

Set up the model-matchmaker hook system in my global Claude Code config. Do exactly the following steps:

────────────────────────────────────────────────────────────
STEP 1 — Create the hooks directory
────────────────────────────────────────────────────────────

Run: mkdir -p ~/.claude/hooks


────────────────────────────────────────────────────────────
STEP 2 — Create ~/.claude/hooks/session-init.sh
────────────────────────────────────────────────────────────

Write this exact content to the file, then run: chmod +x ~/.claude/hooks/session-init.sh

#!/bin/bash
# Session init hook: injects model-tier guidance into every conversation,
# including mandatory rules for sub-agent model selection.
# Adapted from model-matchmaker (https://github.com/coyvalyss1/model-matchmaker)

INPUT=$(cat)