Buzz does not host or bill for models. It routes each agent through an installed harness such as Claude Code or Goose, while the harness connects to the underlying provider. Defaults provide a fallback, and per-agent settings let one shared workspace mix models and harnesses.
The tutorial connects Kimi K3 to Claude Code through Moonshot's Anthropic-compatible endpoint, then assigns it to the main coding agent. A second model reviews the code so that two model families can catch different mistakes.
Lower-stakes agents can use free inference through Nvidia's OpenAI-compatible build API or OpenRouter's free models via Goose. Those tiers are useful for background research and writing, but their rate limits, shared capacity and changing terms make them unsuitable for critical production workloads.
The resulting example uses four agents across three providers: a Kimi K3 coder, a separate Claude reviewer, a free Nvidia-backed researcher and a free OpenRouter-backed writer. The practical advantage is matching model cost and capability to each job without locking the whole team to one vendor.
Watch on YouTube
