quantum-free-router packages a Bifrost-based router that exposes one local OpenAI-compatible endpoint in front of several free-tier LLM providers. It targets long-running coding and research agents, where a single free model hitting a quota, timing out, or being retired should not stop the whole task. The router applies bounded timeout and retry settings and fails over across providers.
An install script sets the router up as a systemd service on Linux or WSL2, and you add free API keys to a JSON config file. A curated model catalog, health-check and certification scripts, and a documented fallback chain come with it, along with a Hermes Agent integration example. Models are sorted into four buckets: active (passed certification), retry-later (quota-limited or transiently slow), configured-only (not yet certified through the router), and quarantine (wrong entitlement, retired, or incompatible).
| Provider | Status |
|---|---|
| OpenCode Zen | Active |
| KiloCode | Active |
| NVIDIA NIM | Active and retry-later mix |
| Mistral, Cerebras, Groq, SambaNova, Cohere | Active when a free key is available |
| Gemini | Retry-later |
| OpenRouter | Quarantine by default |
Features
- Single endpoint: one OpenAI-compatible base URL for any client or agent framework
- Provider failover: routes across OpenCode Zen, KiloCode, NVIDIA NIM, Gemini, Mistral, Cerebras, Groq, SambaNova, and Cohere
- Bounded timeouts and retries: router-level limits so a stalled model does not block the task
- Model certification: scripts that test each model through the router and sort it into active, retry-later, configured-only, or quarantine
- Curated fallback order: a documented chain plus an ordered list for clients that manage their own fallbacks
- Service install: one-command install with systemd setup, health check, and config validation
- Agent examples: Python OpenAI SDK, curl, and Hermes Agent configuration
