Free models on NVIDIA NIM and OpenRouter now hit S-tier on SWE-bench Verified (60%+), but availability is a coin flip. Models get rate-limited, go dark, or spike to 5s+ latency with no notice. I was editing JSON configs by hand every time one died mid-session. frouter is a terminal TUI that pings all free models in parallel every 2 seconds and shows live latency, uptime, and health. You pick one, it writes the config for OpenCode or OpenClaw and launches it. When a model dies, you switch in seconds instead of minutes. How ranking works: models are sorted by availability first, then SWE-bench tier (S+ through C), then rolling average latency. Arena Elo is shown as a secondary signal. All ranking data is bundled and updated with releases; latency and uptime are measured live per session. Some specifics:

Pings are real completion requests, not just health checks Vim-style keybinds, tier filtering (T to cycle S+ → C), search with / frouter --best prints the best model ID to stdout for scripting/CI First-run wizard handles API keys, configs backed up before overwrite Two providers currently: NVIDIA NIM (nvapi- keys) and OpenRouter (sk-or- keys), both free tier

Install and run: npx frouter-cli MIT licensed, TypeScript, 59 commits. Would appreciate feedback on model coverage, provider support, or ranking methodology — especially if you're using free models for agentic coding and have opinions on what's missing. https://github.com/jyoung105/frouter