goose-orchestrator

A multi-model orchestrator-worker MCP extension for Goose. Routes prompts to specialized worker models with VRAM-aware hot-swapping — only the orchestrator and active worker are loaded at any time.

• Automatic task routing — orchestrator LLM decomposes prompts into a dependency graph of sub-tasks, each assigned to a specialized worker role
• VRAM-aware model pool — LRU eviction ensures local models stay within a configurable memory budget; API models bypass the pool
• Speculative preloading — the router predicts the next worker needed and begins loading it during the current step
• Parallel execution — independent steps run concurrently up to a configurable worker limit
• Per-worker temperature — each role has its own sampling temperature (e.g. 0.1 for math, 0.9 for creative)
• Goose-native configuration — all settings managed through MCP tools in Goose's chat UI
• Multi-provider — supports Ollama (local), OpenAI-compatible APIs, and Anthropic

| Role | Default Model | Temperature | Optimized For | |------|--------------|-------------|---------------| | deep_research | qwen3:32b | 0.4 | Multi-hop web research, paper analysis | | local_rag | qwen3:8b | 0.2 | Local document retrieval, code search | | code_gen | qwen2.5-coder:32b | 0.3 | Code generation, refactoring, debugging | | summarizer | qwen3:8b | 0.5 | Condensing outputs, report writing | | math_reasoning | qwen3:32b | 0.1 | Formal proofs, calculations, logic | | creative | llama3:70b | 0.9 | Creative writing, brainstorming |

All roles are fully customizable — change models, add new roles, or remove existing ones.

• Python 3.12+
• uv package manager
• Ollama (for local models) and/or API keys for OpenAI/Anthropic

Clone the repo and