Configuring Multi-Model AI Teams in Buzz

AICodeKing2026-07-29go watch the original →

the gist

Buzz enables per-agent model configuration, allowing users to mix expensive frontier models for coding with free tier models from NVIDIA Build and OpenRouter for research and writing tasks.

Per-Agent Model Orchestration

Buzz functions as an interface for agent harnesses like Claude Code, Codex, and Goose, rather than hosting models itself. Users can assign specific models and harnesses to individual agents, enabling a heterogeneous AI team where model capability is matched to task complexity. This configuration is managed at the agent level, allowing a single workspace to coordinate agents running on different providers simultaneously.

Integrating Kimi K3 for Coding

Kimi K3 can be used as a primary coding model via the Claude Code harness due to its Anthropic-compatible API. To configure this, set the following environment variables in the Claude Code configuration:

ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
ANTHROPIC_AUTH_TOKEN=<your_moonshot_api_key>
ANTHROPIC_MODEL=Kimi-K3

It is recommended to set the auto-compact window to 1,000,000 tokens to leverage the model's full context and disable tool search, as it is incompatible with the Moonshot endpoint.

Leveraging Free API Tiers

Users can reduce costs by routing background agents (research, writing, summarization) through free API providers using the Goose harness.

NVIDIA Build: Provides free access to models like GLM-5.2, Nemotron-3, and DeepSeek-V4. Configure the Goose harness with the base URL https://integrate.api.nvidia.com/v1 and an API key generated from the NVIDIA Build platform.
OpenRouter: Offers a variety of free-tier models, including Nemotron-3 Ultra. Connect via the Goose harness by adding an OpenRouter API key as a provider.

When using free tiers, users should limit agent parallelism to 1 to avoid hitting rate limits, which are typically 20 to 40 requests per minute depending on the provider.