Routed: Open-Source AI Skill Router Released

Routed, an open-source tool, has been released to address the inefficiencies of including all skill definitions in system prompts or using LLMs as routers in AI agent workflows. This tool runs in a 100% local CPU environment and routes skills within 20ms without network transmission.

Hybrid Scoring Pipeline

Routed indexes installed skill directories and evaluates prompts using the following four-stage hybrid method.

• Dense Vector Embeddings (60%): Runs quantized ONNX models such as Arctic Embed S or MiniLM locally on CPU
• Lexical BM25 (25%): Evaluates strict keyword relevance using Okapi BM25
• Exact / Alias Match (10%): Direct command and alias matching
• Metadata (5%): Recency and usage frequency heuristics

Key Features and Compatibility

The v1.1.0 update added support for MCP (Model Context Protocol) servers and native multilingual understanding for over 100 languages. In particular, in MCP mode, instead of loading more than 20 tool schemas into the GPU context window, it exposes only a single route_skill tool and injects the exact schema only when needed to optimize resource usage.

It automatically integrates with major AI development environments such as Cursor, Claude Code, LM Studio, Ollama, and Windsurf, and is released under the MIT license.