Switchboard — ruban-24

Watch out

macOS or Linux only for this first release, and Node.js 22.18 or newer. It uses your coding agent's existing login and never installs the agent for you. The cost and latency table covers classifier requests only, including reporting tags, and the README says actual latency and cost vary; the selected model still does the work and is billed by your own provider. Confidence describes how decisive the classification is, not the probability that the chosen model completes the task correctly, and it feeds separate policy thresholds for capability and selected-model effort that both default to 0.70, while task-type and context confidence go to diagnostics only. Prompt length alone does not determine the route: familiarity, remaining uncertainty, interacting constraints and analysis needed are criteria for the judgments, not separate numeric scores. Low capability confidence applies a balanced minimum rather than dropping to the cheapest tier, and a stronger proposed tier stays stronger. The self-hosted Laya path is experimental: the tested setup uses Laya 0.3.20 with the English checkpoint, tasks longer than 700 characters fall back to uncertain for a new conversation, and the README states that routing quality against Jev has not yet been evaluated. There is no Switchboard account or subscription, and hosted classifier plus native-provider usage are billed separately. Setup makes no AI calls. From a checkout, npm run switchboard also loads .env.local and those values override saved connection settings, whereas installed commands do not load project environment files. If a routed model is missing, or Claude reports that credits are required for Fable, or Claude Code is older than 2.1.280 for Opus 5.5, doctor walks you through the fix. Our README copy is truncated mid-list in the policy section, so fallback behaviour, effort caps and the remaining customization options are not documented here. Copy reviewed against the repository README on 2026-10-04, not from running it.