What it does

Duel Agents sits between you and the LLM APIs you already use. You install a plugin or SDK, swap your base URL to duelagents.com/v1, and pay for a Duel API key. Behind the scenes, it farms your prompt out to multiple models, then serves you back the cheapest response that clears its quality bar. The repo is the integration package: CLI installer, TypeScript SDK, and plugins for Claude Code, Cursor, Codex CLI, and OpenClaw.

The interesting bit

The model-agnostic SDK is the quiet workhorse. It speaks both OpenAI and Anthropic shapes, so you can drop it into existing code without rewriting your chat-completion calls. The “duel-auto” model is a black box — you don’t pick the provider, you just get a price-optimized answer. That abstraction is either liberating or unnerving, depending on your debugging patience.

Key highlights

• OpenAI-compatible and Anthropic-compatible SDK in one package (@duel-agents/sdk)
• One-line installers for Claude Code, Cursor, Codex CLI, and OpenClaw via npx @duel-agents/install
• Requires a Duel API key (duel_<prefix>_<secret>) from their dashboard; no raw Anthropic or OpenAI keys accepted
• Cursor integration needs manual model override in Settings after install
• OpenClaw gets patched via config file; other channels (Telegram, Discord) stay untouched

Caveats

• The actual routing logic, pricing model, and quality thresholds are server-side black boxes; the README doesn’t explain how “cheapest answer that still wins” is determined
• You are locked into Duel’s API key and billing; this is a proxy service, not a self-hosted router
• Cursor setup is two-step (install + manual URL override), and the troubleshooting section suggests skill-copy failures happen enough to warrant a FAQ entry

Verdict

Worth a look if you’re burning through API credits and trust an opaque broker to optimize costs. Skip it if you need provider transparency, predictable latency, or want to keep your existing API keys and direct billing relationships.