smart-model-router

smart-model-router

Auto model-switching plugin for Claude Code — routes prompts to the right tier to save API tokens.

What it does

Every prompt you type is silently classified and compared against your current model:

• git commit all changes → recommends haiku (cheap, fast)
• implement the auth API endpoint → recommends sonnet (default)
• redesign the entire caching architecture → recommends opus (deep reasoning)

No API calls. Zero latency. Fully configurable.

Install via plugin (recommended)

claude plugin marketplace add maiha28781-cloud/smart-model-router
claude plugin install smart-model-router@maiha28781-cloud

Restart Claude Code to activate.

Manual install

git clone https://github.com/maiha28781-cloud/smart-model-router.git
cd smart-model-router
bash install.sh

Configuration

Create ~/.claude/smart-model-router.json (global) or .claude/smart-model-router.json (project-level, overrides global):

{
  "action": "warn",
  "tiers": [
    {
      "name": "haiku",
      "models": ["haiku"],
      "switch_to": "haiku",
      "priority": 1,
      "description": "Simple mechanical tasks",
      "max_word_count": 60,
      "patterns": ["\\bgit\\s+(status|log|diff)\\b"]
    },
    {
      "name": "sonnet",
      "models": ["sonnet"],
      "switch_to": "sonnet",
      "priority": 2,
      "description": "Standard implementation",
      "is_default": true,
      "patterns": ["\\bimplement\\b", "\\bfix\\b"]
    },
    {
      "name": "opus",
      "models": ["opus"],
      "switch_to": "opus",
      "priority": 3,
      "description": "Deep analysis",
      "force_min_word_count": 200,
      "keywords": ["architecture", "analyze", "strategy"]
    }
  ]
}

Add a custom model tier

{
  "name": "my-model",
  "models": ["my-model-id"],
  "switch_to": "my-model-id",
  "priority": 4,
  "description": "My specialized model",
  "keywords": ["ultra complex", "full codebase audit"],
  "patterns": []
}

Extend or replace default keywords/patterns per tier

{
  "opus": {
    "mode": "extend",
    "keywords": ["my-custom-signal"],
    "remove_keywords": ["rethink"]
  },
  "haiku": {
    "mode": "replace",
    "patterns": ["\\bgit status\\b"]
  }
}

mode: extend (default) — adds to built-in keywords/patterns, supports remove_keywords / remove_patterns
mode: replace — discards built-ins entirely, uses only your list

Config options

Field	Type	Default	Description
action	warn | autoswitch	warn	warn = recommend only; autoswitch = change settings.json automatically
default_model	string	sonnet	Assumed model when settings.json has no model field
tiers	array	see defaults	List of model tiers

Tier options

Field	Type	Description
name	string	Display name
models	string[]	Substrings matched against current model
switch_to	string	Value written to settings.json on autoswitch
priority	int	Tier weight (lower = lighter). Must be unique.
mode	extend | replace	Whether to extend or replace default keywords/patterns
keywords	string[]	Case-insensitive keywords matched against prompt
remove_keywords	string[]	Remove specific defaults (extend mode only)
patterns	string[]	Python regex patterns matched against prompt
remove_patterns	string[]	Remove specific default patterns (extend mode only)
max_word_count	int	Only match if prompt is shorter than this
min_word_count	int	Only match if prompt is longer than this
force_min_word_count	int	Force this tier if prompt exceeds N words
force_question_word_count	int	Force this tier if prompt is a long question

Override

Prefix any prompt with ~ to skip classification entirely:

~ do this with whatever model i'm on

Analytics

All routing decisions are logged as structured NDJSON at ~/.claude/logs/smart-model-router.ndjson.

# All-time summary
bash ~/.claude/plugins/cache/maiha28781-cloud/smart-model-router/*/plugins/smart-model-router/hooks/analytics.sh

# Last 7 days
bash analytics.sh --days 7

# Machine-readable JSON
bash analytics.sh --json

Example output:

==================================================
  SMART MODEL ROUTER ANALYTICS
==================================================
  Total prompts classified : 142
  Blocked (wrong model)    : 38 (26.8%)
  Allowed (correct model)  : 97 (68.3%)
  Overridden (~)           : 7 (4.9%)

  Recommendation breakdown:
    sonnet          82  (57.7%)
    haiku           41  (28.9%)
    opus            19  (13.4%)

  Top mismatch flows (blocked):
    sonnet → haiku                  22x
    sonnet → opus                   12x
    opus → sonnet                    4x
==================================================