A proxy that lets Claude and GPT share a house without fighting

What it does

ccflare sits between your code and the Anthropic/OpenAI APIs. You point SDKs at prefixed routes—/v1/anthropic/* or /v1/openai/*—and it forwards requests natively, strips the prefix once, and retries on rate limits using a pool of accounts. There’s also a compatibility layer (/v1/ccflare/*) where model names like anthropic/claude-sonnet-4 or openai/gpt-5.4 pick the provider family for you.

The interesting bit

Most proxies translate schemas; ccflare deliberately doesn’t. Anthropic payloads stay Anthropic, OpenAI stays OpenAI. The value is in the plumbing: per-account rate-limit tracking, automatic failover, a live request stream, and a dashboard that shows history and analytics without you building it yourself. It also handles OAuth for Claude Code and Codex accounts through a TUI, not just static API keys.

Key highlights

• Native passthrough—no payload translation, no schema mangling
• Account-level failover and load balancing with configurable strategies
• Built-in dashboard, TUI, request history, analytics, and health endpoints
• OAuth support for Claude Code and Codex CLI accounts
• Management API for accounts, config, retention, and live log streaming

Caveats

• Requires Bun ≥ 1.2.8; no Node.js runtime mentioned
• README doesn’t specify how rate-limit state is shared (memory? SQLite?)
• Compatibility route provider selection logic (prefers codex, then openai) is present but the exact fallback rules are terse

Verdict

Worth a look if you’re juggling multiple API keys, hitting rate limits, or tired of building internal dashboards for LLM usage. Skip it if you only have one account and one provider; the overhead won’t pay off.