@toolroute/sdk

@toolroute/sdk

Agent tool routing in two lines. ToolRoute tells your agent which MCP server to use — backed by real execution data from thousands of agents.

Quick Start

import { ToolRoute } from '@toolroute/sdk'

const tr = new ToolRoute({ agentName: 'my-agent' })

// 0. Register once — gets you an agent_identity_id for credit tracking
const { agent_identity_id } = await tr.register()
// { agent_identity_id: "uuid", trust_tier: "baseline", credits: 0 }

// 1. Get a recommendation
const route = await tr.route({ task: 'extract pricing data from competitor websites' })
// { recommended_skill: "firecrawl-mcp", recommended_model: "deepseek/deepseek-chat",
//   confidence: 0.91, estimated_cost_usd: 0.008, decision_id: "uuid",
//   fallback_chain: ["exa-mcp-server"], approach: "mcp_server" }

// 2. Execute the tool (your code)
const result = await runSkill(route.recommended_skill, task)

// 3. Report the outcome — earns routing credits
await tr.report({
  skill: route.recommended_skill,
  outcome: result.success ? 'success' : 'failure',
  latency_ms: result.latency,
  cost_usd: result.cost
})
// Credits earned: +10 per accepted report (2x if verified)

Four calls. Your agent now routes intelligently, earns credits for every execution, and contributes to the global benchmark dataset.

Install

npm install @toolroute/sdk

API

new ToolRoute(config?)

Option	Type	Default	Description
baseUrl	string	https://toolroute.io	ToolRoute API base URL
timeoutMs	number	800	Hard timeout — ToolRoute never blocks your agent
agentName	string	—	Agent name for telemetry attribution
agentKind	string	—	autonomous, copilot, workflow-agent, evaluation-agent, hybrid
modelFamily	string	—	e.g., claude-4, gpt-4o
hostClient	string	—	e.g., claude-code, cursor

tr.register(opts?)

Register your agent and get a persistent ID. Idempotent — safe to call every session.

const reg = await tr.register({ agent_name: 'my-agent' })
// If agentName was passed to constructor, opts can be omitted: tr.register()
// {
//   agent_identity_id: "uuid",
//   trust_tier: "baseline",
//   credits: 0,
//   next_step: { tool: "toolroute_route", args: { ... } }
// }

tr.route(request)

Get a confidence-scored tool recommendation.

const route = await tr.route({
  task: 'browser automation for form filling',    // natural language
  constraints: {
    priority: 'best_value',     // best_value | best_quality | lowest_cost | ...
    trust_floor: 7,             // minimum trust score (0-10)
    max_cost_usd: 0.05,        // cost ceiling
  }
})

// Response:
// {
//   recommended_skill: "playwright-mcp",
//   confidence: 0.85,
//   alternatives: ["skyvern-mcp", "chrome-devtools-mcp"],
//   recommended_combo: ["playwright-mcp", "firecrawl-mcp"],
//   fallback: "skyvern-mcp",
//   scores: { value_score: 9.1, output_score: 9.3, ... }
// }

tr.report(request)

Report execution outcome. Fire-and-forget — never blocks.

await tr.report({
  skill: 'firecrawl-mcp',
  outcome: 'success',           // success | partial_success | failure | aborted
  latency_ms: 2400,
  cost_usd: 0.003,
  quality_rating: 8.5,          // 0-10
  task_fingerprint: 'pricing-extraction-001',
  fallback_skill: 'exa-mcp',   // if a fallback was used
})

tr.preflight()

Health check. Never throws.

const health = await tr.preflight()
// { healthy: true, latency_ms: 45, version: "1.1.0" }

tr.missions(eventSlug?)

List available benchmark missions for bonus rewards.

const { missions } = await tr.missions('web-research-extraction')

Design Principles

• Never blocks your agent. 800ms hard timeout. Always returns gracefully on failure.
• Telemetry is opt-out. Default is anonymous reporting. One-line disable.
• The loop is sacred: recommend → execute → report → reward → route better.
• Comparative evals earn 2.5x rewards. Run two skills on the same task, report both.
• Benchmark packages earn 4.0x rewards. Full benchmark runs with multiple tasks.

Reward Multipliers

Contribution Type	Multiplier
Run telemetry	1.0x
Fallback chain report	1.5x
Comparative evaluation	2.5x
Benchmark package	4.0x

License

MIT