ClaudeRouter

Stop burning Opus tokens on grep. ClaudeRouter classifies every prompt before it hits the model — routing simple tasks to Haiku, feature work to Sonnet, and architecture to Opus. Ships as a Claude Code plugin. Zero config.

Prerequisites

• Node.js 18+
• Claude Code installed (install guide)
• Claude Pro or Max subscription (required for subagent delegation)
• jq — the hook script depends on it (brew install jq / apt install jq)

Installation

npm install -g @0dust/claude-router
claude-router init

From source

git clone https://github.com/0dust/ClaudeRouter.git
cd ClaudeRouter
npm install
npm run build
npm install -g .
claude-router init

This will:

• Verify dependencies (jq)
• Register the UserPromptSubmit hook in ~/.claude/settings.json
• Append the routing directive to your project's CLAUDE.md (with markers for clean removal)

You can also target a specific project directory:

claude-router init /path/to/your/project

Verify It's Working

Run the diagnostic check:

claude-router doctor

After a few prompts, check your routing stats:

claude-router stats

If counts are incrementing, routing is active. You can also check the hook is registered:

cat ~/.claude/settings.json | jq '.hooks.UserPromptSubmit'

How It Works

ClaudeRouter intercepts every UserPromptSubmit hook, classifies the prompt's complexity, and injects a routing directive into Claude's context. Claude then delegates to the appropriate subagent model.

Tier	Model	Example Prompts	Savings vs Opus
LOW	Haiku	"what does this function do?", "yes do it", file reads, grep	~95%
MEDIUM	Sonnet	"add an endpoint for X", "fix this bug", "write tests for Y"	~60%
HIGH	Opus	"redesign the auth layer", "review this architecture"	0% (correct spend)

Classification Pipeline

1. Pre-Haiku heuristics (synchronous, zero cost) — catches ~30% of LOW prompts via token count, keyword matching, and confirmation detection. No API call needed.
2. Haiku classification (async, <$0.001) — sends the prompt to Haiku with a structured classification prompt. Returns LOW, MEDIUM, or HIGH.
3. Context injection — the routing directive is injected as context. Claude reads it and delegates to the appropriate subagent.

Routing Directives

• LOW → Claude spawns a Haiku subagent via the Agent tool
• MEDIUM → Claude spawns a Sonnet subagent via the Agent tool
• HIGH → Claude handles the task directly with full reasoning
• Override → User prefixed with //opus, Claude handles directly

Stats

Track your routing efficiency:

claude-router stats

ClaudeRouter — last 7 days
─────────────────────────────────────────
Prompts routed:        847
LOW  → Haiku:          312   (36.8%)
MED  → Sonnet:         431   (50.9%)
HIGH → Opus:           104   (12.3%)
Estimated Opus saved:   148.6K tokens
Follow-up rate (LOW):   4.1%    ← routing accuracy
Manual overrides:       7
─────────────────────────────────────────

Use --days N to change the window: claude-router stats --days 30

Failure Behavior

ClaudeRouter is designed to never block Claude Code:

• Classifier fails: Falls back to MEDIUM (Sonnet) — the safe middle ground
• API timeout: Haiku classification has a 3-second timeout; on timeout, falls back to MEDIUM
• No internet: Pre-Haiku heuristics still work (catches ~30% of prompts); the rest fall back to MEDIUM
• Missing dependencies: Hook exits silently with no directive; Claude handles the prompt normally on whatever model the session is using
• Any unexpected error: The hook always exits 0 and never blocks the user's prompt

Configuration

ClaudeRouter works with zero configuration. To customize, create .claude-router.json in your project root or ~/.claude-router.json globally:

{
  "tiers": {
    "LOW": "claude-haiku-4-5-20251001",
    "MEDIUM": "claude-sonnet-4-6",
    "HIGH": "claude-opus-4-6"
  },
  "fallback": "claude-sonnet-4-6",
  "conservative": false,
  "override_keyword": "//opus"
}

Field	Default	Description
tiers.LOW	claude-haiku-4-5-20251001	Model for trivial tasks
tiers.MEDIUM	claude-sonnet-4-6	Model for standard engineering work
tiers.HIGH	claude-opus-4-6	Model for deep reasoning tasks
fallback	claude-sonnet-4-6	Model used when classification fails
conservative	false	Shift all routes one tier up (LOW→Sonnet, MEDIUM→Opus)
override_keyword	//opus	Prefix to force Opus on any turn

Config is merged in order: hardcoded defaults → ~/.claude-router.json → CWD .claude-router.json. Later values override earlier ones.

Team Configuration

To share routing config across a team, commit .claude-router.json to your repo (you may need to remove it from .gitignore). Individual developers can still override with ~/.claude-router.json for personal preferences.

Troubleshooting

Stats show zero events after using Claude Code The hook may not be registered. Check:

cat ~/.claude/settings.json | jq '.hooks.UserPromptSubmit'

If empty, re-run claude-router init.

Hook not firing

1. Verify jq is installed: command -v jq
2. Verify the hook script exists at the path shown in settings.json
3. Verify the hook is executable: ls -la $(which claude-router)

jq not installed The hook exits silently without jq. Install it:

• macOS: brew install jq
• Ubuntu/Debian: sudo apt install jq
• Arch: sudo pacman -S jq

claude-router command not found after install If installed via npm install -g ., ensure your npm global bin directory is in your PATH: npm bin -g

I see [ROUTER] lines in my Claude Code transcript This is normal. The routing directive is visible in the transcript as hook output, but Claude follows it silently and does not mention it in responses.

All prompts routing to Sonnet (MEDIUM) This is the fallback behavior when Haiku classification fails. Verify that Claude Code can reach the Anthropic API (this requires an active Claude Pro or Max subscription).

Override Keyword

Prefix any prompt with //opus to bypass classification and force Opus:

//opus explain the tradeoffs between these two architectures

This routes directly to Opus regardless of what the classifier would have chosen. The keyword is stripped from the prompt before processing.

SDK Usage

The routing logic is available as a standalone SDK for any Claude-based agent stack:

import { classify, route, createRouter } from 'claude-router';

// Single classification call
const result = await classify('fix the typo on line 42');
// { tier: 'LOW', source: 'signal', latency_ms: 0, prompt_tokens: 7 }

// Full routing decision
const config = loadConfig();
const decision = await route('add user authentication', config);
// { model: 'claude-sonnet-4-6', tier: 'MEDIUM', source: 'haiku', ... }

// Stateful router with telemetry
const router = createRouter({ telemetry: true });
const d = await router.route('redesign the auth layer');
console.log(router.stats());
// { total: 1, low: 0, medium: 0, high: 1, overrides: 0, avg_latency_ms: 12 }

API

• classify(prompt) — Returns ClassificationResult with tier, source, latency, and token count
• quickClassify(prompt) — Synchronous heuristic-only classification. Returns tier or null
• route(prompt, config) — Full routing decision with model, directive, and metadata
• loadConfig() — Load merged config from defaults + global + local files
• createRouter(options?) — Stateful router instance with .route() and .stats() methods

How the Plugin Works

ClaudeRouter uses Claude Code's UserPromptSubmit hook to inject routing directives as context. The CLAUDE.md file in your project root instructs Claude to follow these directives transparently — delegating to Haiku or Sonnet subagents for lower-complexity tasks.

This approach:

• Requires no model mutation — works within the existing hook API
• Is minimally visible — routing directives appear in the transcript but Claude does not mention or acknowledge them to the user
• Includes an infinite loop guard — subagents don't re-trigger the classification hook

Uninstallation

claude-router remove
npm uninstall -g claude-router

This removes:

• The UserPromptSubmit hook from ~/.claude/settings.json
• The <!-- claude-router:start --> ... <!-- claude-router:end --> block from your project's CLAUDE.md

Telemetry data in ~/.claude-router/ is preserved. Delete it manually if desired.

Contributing

See CONTRIBUTING.md.

License

MIT