Routed

What bshea-1/Routed does

bshea-1/Routed indexes Agent Skills available to coding environments and selects the most relevant skill for each coding prompt. Its routing engine combines exact and alias matches, Okapi BM25 lexical retrieval, and dense semantic embeddings. It can also decompose compound prompts and activate more than one skill.

The project is intended for local use across macOS, Linux, and Windows. Prompt routing happens on the machine, so the documented workflow does not require a network request or an LLM call for routing. The README reports sub-20ms local search latency, but actual performance depends on the machine and indexed skills.

How it works

A prompt passes through a hybrid scoring pipeline. Exact matches contribute direct evidence, BM25 handles lexical overlap and terminology, and dense embeddings cover semantic similarity. Metadata such as routing history and decay can also participate in scoring. The engine ranks candidate skills and applies confidence handling before returning a selection.

Routed includes separate behavior for anchored and unanchored queries. Queries with lexical or exact matches use a lower confidence threshold, while queries relying only on semantic similarity use a stricter threshold. The README also describes handling for negated intents, framework-specific penalties, multilingual prompts, and no-skill inputs.

The routed tune --folds 5 --apply command evaluates candidate scoring parameters with stratified cross-validation and stores selected weights in the local SQLite index.db. routed tune --reset restores the baseline configuration. The repository includes a local benchmark command for measuring routing accuracy, recall, and no-skill behavior.

Setup and configuration

bshea-1/Routed provides standalone installers named RoutedSetup.exe for Windows, RoutedSetup.pkg for macOS, and RoutedSetup.deb for Linux through GitHub Releases. The documented source setup clones the repository, runs npm install, builds with npm run build, and can then run the benchmark with npm run benchmark.

Start the MCP interface with routed mcp. The README identifies LM Studio, Cursor, Claude Desktop, Windsurf, Continue, and Cline as MCP environments where this mode can reduce routing-related context. The provided material does not specify a required configuration file, MCP transport option, or environment variable.

Tools and capabilities

• Route prompts to one or more relevant Agent Skills with routed route.
• Expose routing through MCP with routed mcp.
• Tune scoring weights using routed tune, with optional reset support.
• Run local evaluation through routed benchmark and request JSON output with --json.
• Check and repair host adapters with routed doctor --fix.
• Check for and apply updates with routed update.
• Search across multilingual prompts, including German, Spanish, French, Japanese, and other languages described by the project.

Limitations and notes

Routing depends on the skills that have been scanned and indexed; the provided material does not define a fixed catalog of skills. The system is designed for coding prompts and Agent Skill selection, rather than general-purpose task routing.

Reported accuracy, recall, confidence, and false-activation figures come from the repository's stated evaluation sets and specific releases, including v1.6.5 and v1.6.96. They should be treated as project benchmarks, not guarantees for every skill collection or workload. Local CPU execution avoids remote prompt routing, but embedding and indexing behavior still depends on the installed application and local data.