@dalek-ai/router

TypeScript SDK for agent-tool-router — picks the right tools for an agent task from a catalog of 18 000.

The default endpoint is a hosted API at dalek-ai-router-api.hf.space. No API key, no signup.

Install

npm install @dalek-ai/router

Requires Node 18+ (uses the built-in fetch).

Use

import { route } from "@dalek-ai/router";

const result = await route({
  task: "annule ma commande et rembourse-moi",
  k: 3,
});

console.log(result.tools);
// [
//   { name: "get_order_details", score: 0.26 },
//   { name: "get_user_details", score: 0.25 },
//   { name: "return_delivered_order_items", score: 0.24 }
// ]
console.log(result.model);      // "baseline-v1-desc-hybrid-multilingual-next-v1"
console.log(result.latency_ms); // ~200

History-aware routing

Pass the previous tools called to bias the next-tool prediction (uses a Markov-2 history bigram with stupid-backoff under the hood):

const result = await route({
  task: "rembourse-moi maintenant",
  history: ["get_order_details", "verify_refund_eligibility"],
  k: 3,
});

Reusable client

import { Router } from "@dalek-ai/router";

const r = new Router(); // defaults to https://dalek-ai-router-api.hf.space
const top3 = await r.route("send slack to my team");

// or point at a self-hosted instance
const self = new Router("https://router.mycompany.internal");

Errors

Failed requests throw RouterError with .status and .body:

import { RouterError } from "@dalek-ai/router";

try {
  await route({ task: "..." });
} catch (e) {
  if (e instanceof RouterError) console.error(e.status, e.body);
}

Hosted API specs

• URL: https://dalek-ai-router-api.hf.space
• Median latency: ~200 ms (shared free CPU on Hugging Face Spaces)
• Default model: baseline-v1-desc-hybrid-multilingual-next-v1 (FR + EN, Pareto-dominates 3 LOSO benchmarks)
• Rate limits: HF Spaces public quotas (generous, not for production scale)
• Higher limits / private deployments: drop your handle on the Waitlist discussion

Numbers

• 18 671 tools indexed (Hermes function-calling, ToolACE, tau-bench)
• 14 000 agent traces in the training corpus
• 77.4% top-3 next-tool accuracy (Markov-2 backoff on the multilingual encoder)
• 100% top-3 on tau-bench position 2 (228 test calls, zero errors at this horizon)
• 9 ms p50 latency in-process (local MacBook MPS) with the Python SDK

Full benchmark numbers, eval scripts, and the dataset rebuild script: github.com/dalek-ai/agent-tool-router

License

MIT.