LLM Providers

Honeybee supports multiple LLM providers. Mix providers within a single brood — use fast inference for simple tasks and capable models for complex reasoning.

Supported providers

Provider	Alias	Default Model	Best for
Cerebras	fast	llama-3.3-70b	Fast inference, low latency
Groq	—	llama-3.3-70b	Fast inference, OpenAI-compatible
Anthropic	smart	claude-sonnet	Complex reasoning, tool use
OpenAI	—	gpt-4o	Broad ecosystem, vision
Ollama	local	llama3.3	Local inference, no API costs

Setting provider keys

Local development

Provider keys are sourced from ~/.secrets/*.env when running wgl up:

~/.secrets/cerebras.env

CEREBRAS_API_KEY=csk-...

# ~/.secrets/groq.env
GROQ_API_KEY=gsk-...

# ~/.secrets/anthropic.env
ANTHROPIC_API_KEY=sk-ant-...

# ~/.secrets/openai.env
OPENAI_API_KEY=sk-...

Cloud (Colony)

# Set a user-level key (used by all hives)
wgl secret set cerebras

# Set a hive-specific key
wgl secret set cerebras --hive my-project

# List stored keys
wgl secret list

Configuration in brood.yaml

Default provider

# Full format
provider: cerebras/llama-3.3-70b

# Alias
provider: fast

Per-agent override

hives:
  main:
    agents:
      - role: architect
        provider: anthropic/claude-opus    # Complex decisions
      - role: developer
        provider: cerebras/llama-3.3-70b   # Fast execution
      - role: reviewer
        provider: fast                     # Alias for cerebras

Provider details

Cerebras

Fastest inference available. TCP warming reduces time-to-first-token.

provider: cerebras/llama-3.3-70b
# or
provider: cerebras/llama-3.1-8b

• TCP warming enabled by default in incubator (long-running)
• Disabled in CF Workers (warmTCPConnection: false)
• 2000+ tokens/second for streaming

Groq

Fast inference with OpenAI-compatible API.

provider: groq/llama-3.3-70b

• OpenAI SDK fork (identical API surface)
• CF Workers: requires nodejs_compat flag

Anthropic

Most capable models. Best for Queens and complex reasoning.

provider: anthropic/claude-opus
provider: anthropic/claude-sonnet
provider: anthropic/claude-haiku

• Native tool use support
• Vision capabilities
• Input caching for repeated system prompts

OpenAI

Broad model selection and ecosystem.

provider: openai/gpt-4o
provider: openai/gpt-4o-mini

Ollama (local)

Run models locally with zero API costs. Requires Ollama running on the network.

provider: ollama/llama3.3
provider: ollama/qwen2.5:72b
provider: ollama/deepseek-r1:70b

Default endpoint: http://localhost:11434. Override with OLLAMA_HOST environment variable.

Cost optimization patterns

# Cheap: drone + fast provider for review/voting
- role: reviewer
  type: drone
  provider: fast

# Medium: worker + mid-tier for implementation
- role: developer
  type: worker
  provider: cerebras/llama-3.3-70b

# Premium: claude type + opus for architecture
- role: architect
  type: claude
  provider: anthropic/claude-opus

# Free: local ollama for development/testing
- role: tester
  type: worker
  provider: local

Use the cheapest provider that meets the task requirements. Simple CSS → Haiku. API design → Sonnet. Architecture → Opus.