ToolRoute

Routing layer for AI agents. One call returns the best MCP server and LLM for any task — scored on 132 real benchmark executions.

Quick start

Add to any MCP client (Claude Code, Cursor, Windsurf, Cline):

config.json

{
  "mcpServers": {
    "toolroute": {
      "url": "https://toolroute.io/api/mcp"
    }
  }
}

Or via HTTP:

Terminal

curl -X POST https://toolroute.io/api/route \
  -H "Content-Type: application/json" \
  -d '{"task": "search the web for recent AI papers"}'

config.json

{
  "approach": "mcp_server",
  "recommended_skill": "exa-mcp-server",
  "recommended_skill_name": "Exa MCP Server",
  "recommended_model": {
    "slug": "claude-haiku-4-5-20251001",
    "display_name": "Claude Haiku 4.5",
    "provider": "anthropic",
    "tier": "cheap_chat",
    "provider_model_id": "anthropic/claude-haiku-4-5-20251001",
    "input_cost_per_mtok": 1.00,
    "output_cost_per_mtok": 5.00
  },
  "confidence": 0.91,
  "alternatives": ["brave-search-mcp", "tavily-mcp"],
  "fallback": "brave-search-mcp"
}

recommended_model is always an object, not a bare string — the inner slug is the canonical model identifier.

How it works

Every task falls into one of three approaches:

Approach	When	Returns
`direct_llm`	Task needs only an LLM (code, writing, analysis)	Best model + cost estimate
`mcp_server`	Task needs an external tool (search, email, calendar)	Best tool + best model
`multi_tool`	Compound task ("send Slack AND update Jira AND email")	Ordered orchestration chain

Routing uses an LLM classifier (~$0.00001/call) for task understanding, then ranks candidates on a 5-dimension score:

Code

Value Score = 0.35 × Output Quality
            + 0.25 × Reliability
            + 0.15 × Efficiency
            + 0.15 × Cost
            + 0.10 × Trust

Every reported outcome updates the scores. The routing gets more accurate as more agents use it.

Benchmark results

132 blind A/B executions across code, writing, analysis, structured output, and translation.

	ToolRoute	Fixed GPT-4o
Quality wins	6	0
Ties	9	9
Losses	0	—
Avg cost	$0.001–0.01	$0.03–0.10

API

Endpoint	Method	Description
`/api/route`	POST	Route a task to best MCP server + LLM (unified)
`/api/route/model`	POST	Route to best LLM model only (no MCP server)
`/api/mcp`	POST (JSON-RPC)	MCP server — 16 tools
`/api/mcp`	GET (SSE)	SSE transport for MCP clients
`/api/report`	POST	Report MCP server outcome (lightweight)
`/api/contributions`	POST	Advanced MCP skill telemetry (requires `skill_id` or `skill_slug` in payload)
`/api/report/model`	POST	Report LLM model outcome — use this for model telemetry, not `/api/contributions`
`/api/verify/model`	POST	Verify model output quality
`/api/skills`	GET	Search MCP server catalog
`/api/agents/register`	POST	Register agent identity
`/api/agents/preferences`	POST	Set routing preferences (Strategy D Phase 2 — `allow_china`, `regulated_industries`)
`/api/health`	GET	Service health check (DB + uptime)
`/api/metrics`	GET	Public aggregate platform metrics (no auth)

Full reference at toolroute.io/api-docs

SDK

Terminal

npm install @toolroute/sdk

server.ts

import { ToolRoute } from '@toolroute/sdk'

const tr = new ToolRoute()
const rec = await tr.route({ task: 'parse this CSV and summarize it' })
// execute with rec.recommended_model ...
await tr.report({ skill: rec.recommended_skill, outcome: 'success', latency_ms: 1400 })

Self-hosting

bash

git clone https://github.com/grossiweb/ToolRoute.git
cd ToolRoute
cp .env.local.example .env.local
npm install
npm run dev

Requires: NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY

How routing works

ToolRoute classifies each task using an LLM classifier (Gemini Flash Lite, ~$0.00001/call) with a keyword fallback. The resulting tier maps to a specific model via src/lib/routing/tiers.ts. Live pricing and capability data come from the models table. See docs/architecture.md for the full picture.

Stack

Next.js 14 (App Router) · Supabase (Postgres) · Vercel

License

MIT