CCR Rust

CCR-Rust

CCR-Rust is a lightweight, self-hosted LLM router that sits between your AI coding tools (Claude Code, Codex, OpenCode, or any OpenAI-compatible client) and your model providers. Your client talks to one local URL; CCR-Rust decides which provider and model actually handles each request, translates between API formats when needed, and fails over automatically when a provider has problems. It is aimed at developers and teams who want to keep using their favorite coding CLI even when they need to switch providers, spread load across several models, or keep working after a usage limit runs out.

The most common reason people install it: when your Claude plan runs out, keep using Claude Code with GLM, MiniMax, DeepSeek, or another provider instead of changing tools.

Why route at all

A coding agent and its model provider are two independent layers, and the failures that actually kill long agent sessions happen in the provider layer: quota windows (5-hour, weekly, monthly) run out, providers emit 429 storms at peak hours, endpoints have regional outages, models get deprecated or silently rerouted, and pricing or terms change without notice. No agent, however good, can fix a provider that is rate-limiting it. A router places one local, provider-agnostic endpoint between the agent and that entire failure surface:

- Failover instead of failure. Providers are arranged in tiers; 5xx errors and timeouts cascade to the next tier automatically.
- Held retries instead of dropped work. With retry sweeps enabled, a request that exhausts every tier is held in flight and re-cascaded after a cooldown, so the agent observes latency instead of a 429. Deterministic rejections (401/402/403/404) are attempted once and skipped afterward.
- Client choice decoupled from provider…