@toolroute/hook

Auto-route every agent task through ToolRoute. Install once — every task gets the best MCP server and LLM model recommendation, automatically.

Install

npm install @toolroute/hook

Quick Start

import { createHook } from '@toolroute/hook'

const hook = createHook({
  agentName: 'my-agent',
  verbose: true, // logs routing decisions
})

// Register (optional, earns 2x credits)
await hook.register()

// Wrap any task — routing + reporting happens automatically
const result = await hook.run('scrape competitor pricing data', async (routing) => {
  console.log(`Best tool: ${routing.skillName}`) // e.g., "Firecrawl MCP"
  console.log(`Best model: ${routing.model}`)    // e.g., "claude-haiku-4-5-20251001"
  console.log(`Confidence: ${routing.confidence}`)

  // ... your actual task logic here ...
  return { prices: [29, 49, 99] }
})

console.log(result.data)          // { prices: [29, 49, 99] }
console.log(result.executionMs)   // 1847
console.log(result.creditsEarned) // 5

Just Route (no wrapper)

const routing = await hook.route('draft an email to a client')
// routing.skill → "gmail-mcp"
// routing.model → "claude-haiku-4-5-20251001"
// routing.confidence → 0.87

A/B Testing

// Set bypass: true to skip routing — compare routed vs unrouted performance
const hook = createHook({ bypass: Math.random() > 0.5 })

How It Works

1. Before your task — calls POST /api/route with your task description
2. Returns routing context — best skill, model, alternatives, confidence score
3. You execute — use the recommendation (or ignore it)
4. After your task — reports outcome to ToolRoute (fire-and-forget)
5. Earns credits — 3-10 credits per report, 2x if verified

Links

• ToolRoute — Skill intelligence for AI agents
• API Docs
• MCP Server Catalog
• Challenges — Earn 3x credits

Changelog

1.0.2 (2026-05-20)

• Fix: hook.route() returned model: null for every call because it read recommended_model.alias / .recommended_alias — fields that don't exist on the /api/route response. Now reads recommended_model.slug.
• Fix: hook.run() called reportOutcome() twice per task on the success path (once in the finally block, once after). Telemetry was double-counted. Now fires exactly one POST /api/contributions per task on both success and failure paths.

1.0.1

• Initial public release.