Multi-LLM Routing

Overview

Armament supports multiple LLM providers simultaneously. Each channel/agent can be configured with a specific provider and model. The ProviderPool manages provider instances and handles deduplication.

Provider Pool

Providers are configured in ~/.armament/config.yaml:

providers:
  - type: bedrock
    region: us-west-2
    models:
      - name: claude-sonnet-4-20250514
  - type: openai
    apiKey: ${OPENAI_API_KEY}
    models:
      - name: gpt-4o
      - name: gpt-4o-mini
  - type: anthropic
    apiKey: ${ANTHROPIC_API_KEY}
    models:
      - name: claude-sonnet-4-20250514
  - type: gemini
    apiKey: ${GEMINI_API_KEY}

Per-Agent Model Selection

When spawning an agent, specify the provider and model:

/spawn reviewer --model claude-sonnet-4-20250514 --provider anthropic

Or change the model for an existing channel:

/model gpt-4o
/provider openai

Supported Providers

Provider	Auth	Models
Bedrock	AWS SSO / IAM	Claude, Mistral, Llama
Anthropic	API Key	Claude Opus, Sonnet, Haiku
OpenAI	API Key	GPT-4o, GPT-4o-mini, o-series
Gemini	API Key	Gemini Pro, Flash
Ollama	Local	Any local model

Fallback Chain

Configure a fallback chain to try alternative models if the primary is unavailable:

/fallback gpt-4o claude-sonnet-4-20250514 gemini-pro

Rate Limiting

Rate limits are tracked per provider. When a provider returns rate limit errors, the system waits and retries automatically.

Planned Enhancements

Cost-optimized routing: Select cheapest model meeting capability threshold
Quality-first routing: Route complex tasks to best model, simple tasks to cheaper models
Rule-based routing: Match tasks to models based on complexity, type, context size
Observability dashboard: /route stats for model usage breakdown