Engines

Four engine families. One agent definition.

The engine is the brain Auggy is wearing — the model that does the actual reasoning. Auggy ships first-class adapters for hosted and local providers. Pick one, change a line of YAML, redeploy. Same agent, different brain.

Not hidden behind an abstraction layer. Each adapter is ~150 lines of TypeScript that converts the kernel’s tool calls into that provider’s native format.

Anthropic

Native client to the Anthropic Messages API.

Claude with full tool-use parity. Default to Opus 4.6 for the hardest reasoning, Sonnet 4.6 for balance, Haiku 4.5 for speed.

Models

OPUS 4.6SONNET 4.6HAIKU 4.5

agent.yaml

engine:
  provider: anthropic
  model: claude-opus-4-6

Streams via SSE. 1M-context Opus supported.

OpenAI

Chat Completions with full tool-use.

GPT-5 and the OpenAI lineup behind the same kernel contract. Same agent definition, different brain.

Models

GPT-5.5GPT-5.4GPT-5.4-MINI

agent.yaml

engine:
  provider: openai
  model: gpt-5

Function-call → kernel tool bridge handled in the adapter.

OpenRouter

One adapter, hundreds of models.

Route through OpenRouter to reach Qwen, Llama, Mistral, Gemini, DeepSeek, and the entire OR catalog. One API key, model-agnostic.

Models

QWENLLAMAMISTRALANY OR MODEL

agent.yaml

engine:
  provider: openrouter
  model: qwen/qwen-3-72b-instruct

Best for cost optimization, multi-model fallback, exotic models.

Ollama

Local or remote Ollama models.

Run against localhost for private local development, or point at a remote Ollama-compatible host.

Models

QWENLLAMAMISTRALLOCAL MODELS

agent.yaml

engine:
  provider: ollama
  model: qwen3:8b

Local mode needs no API key. Remote bearer auth is optional.

The swap

Engine choice is one line.

Multi-engine support isn’t config theater. The kernel speaks one tool-use protocol; each engine adapter converts to and from the provider’s native format. Your agent code doesn’t change.

A/B test models in production. Fall back when one provider degrades. Use Anthropic for the cautious step, OpenAI for the broad one, OpenRouter for the experimental edge.

Same agent, four engine families

# Pick one. Restart. Done.

engine:
  provider: anthropic
  model: claude-opus-4-6

# engine:
#   provider: openai
#   model: gpt-5

# engine:
#   provider: openrouter
#   model: qwen/qwen3.5-397b-a17b

# engine:
#   provider: ollama
#   model: qwen3:8b

# Everything below stays identical.
augments:
  - { name: identity, type: fileMemory, options: { source: ./identity.md } }
  - { name: web,      type: webTransport, options: { port: 8080 } }

Want to add another?

Engine adapters are small and provider-specific. The contract is in src/engines/ on GitHub.

Quickstart Source on GitHub