@robot-resources/router

Smart model selection for AI agents — pick the cheapest LLM that can handle each prompt.

The router classifies a prompt by task type (coding / reasoning / analysis / simple_qa / creative / general), filters the model catalog by capability, and returns the cheapest qualifying model. Typically saves 60–90% on coding/QA mixes vs always picking the top model.

Two ways to use it

As an OpenClaw plugin (default npx robot-resources install)

The router IS the OC plugin. It registers an in-process HTTP server on 127.0.0.1:18790 inside OC's gateway process, dispatches LLM calls there, runs the keyword classifier, picks a model, and forwards the request to the user's configured provider (currently Anthropic) using the user's existing API key. No daemon, no Python, no system service. Lifetime is tied to the OC gateway — the server starts when OC loads the plugin and dies when OC exits.

npx robot-resources                # installs the plugin into ~/.openclaw/extensions/robot-resources-router/

The unified wizard handles plugin install, scraper MCP registration, and config patching. See robot-resources on npm for the wizard.

As a public JS library (any agent framework)

For agents that don't run inside OpenClaw — LangChain, LangGraph.js, Mastra, etc. — import the routing decision directly:

npm install @robot-resources/router

import { routePrompt } from '@robot-resources/router/routing';

const decision = routePrompt('write a python function that reverses a string');
// {
//   selected_model: 'claude-haiku-4-5',
//   provider: 'anthropic',
//   savings_percent: 68.0,
//   task_type: 'coding',
//   capability_score: 0.86,
//   reasoning: '...',
// }

Pure ESM, zero dependencies. The classifier runs offline on the keyword fast-path (~5ms, covers ~70% of prompts); slow-path classifier needs a network call to Gemini (configurable via asyncRoutePrompt).

import { asyncRoutePrompt } from '@robot-resources/router/routing';

// Async path — same shape, may call Gemini for ambiguous prompts.
const decision = await asyncRoutePrompt(prompt);

The routing module never reads user provider keys — modelsDb and availableProviders are caller parameters. Safe to use in non-OC contexts.

Other consumption paths

What it ships

• index.js — OC plugin entry (sync register shim)
• lib/local-server.js — in-process HTTP server (Anthropic-messages compatible)
• lib/routing/ — classifier + selector + models catalog (importable as @robot-resources/router/routing)
• lib/telemetry.js — opt-in telemetry helper (importable as @robot-resources/router/telemetry)
• lib/self-update.js — daily npm version poll + atomic swap

Public exports

// from package.json `exports`
import x from '@robot-resources/router';            // OC plugin shim (default)
import x from '@robot-resources/router/routing';    // routing API (routePrompt, asyncRoutePrompt, MODELS_DB)
import x from '@robot-resources/router/telemetry';  // createTelemetry factory

Provider support

Today: Anthropic (in-process server forwards to api.anthropic.com). The classifier catalog includes OpenAI + Google + Anthropic models, so the JS library can return decisions for any provider — but the in-process server's request forwarding is single-lab. Multi-lab dispatch is a future expansion that lands in a follow-up.

Configuration (OC plugin path)

The plugin reads the user's Anthropic key from ~/.openclaw/agents/<id>/agent/auth-profiles.json (the OC native auth store). The user never enters a key into Robot Resources directly. We never see, store, or transmit user provider keys.

~/.robot-resources/config.json stores the anonymous Robot Resources api_key used for telemetry — provisioned at install via POST /v1/auth/signup. Optional. Routing works without it.

Telemetry

Fire-and-forget, opt-in. Events emitted from the in-process server: local_server_started, route_completed, route_failed, local_server_no_key, local_server_upstream_failed. Events from the JS library (when telemetry is wired by the consumer): route_via_lib. Disable: don't supply an apiKey.

License

MIT