Skip to content

Guides

Implementation-focused guides for common production patterns.

Core guides

  • Strategy — Gearbox strategy and mode selection
  • Agent guide — Agent lifecycle and configuration
  • Agent config guide — AgentConfig deep dive
  • Autonomous mode — v0.7.14 harness: why the changes existed, what improved, and current examples
  • MCP integration — Model Context Protocol universal adapter — 🟢 nucleusiq-mcp 0.1.1 Stable, nucleusiq>=0.7.12, mcp>=1.28.1 + legacy OpenAI server-side path
  • File handling — Attachment vs Tool vs Both

Provider guides

Provider Package Version Status
OpenAI nucleusiq-openai 0.7.1 🟢 Stable — Chat Completions, Responses API, native tools, usage-accounting fix
OpenAI-compatible nucleusiq-openai-compatible 0.1.0 🟢 New, Stable — vLLM / SGLang / llama.cpp / LM Studio / Azure OpenAI v1 / any Chat Completions server. Requires nucleusiq>=0.7.13
Gemini nucleusiq-gemini 0.3.1 🟢 Stable — google_search / code_execution server-tool observability, thinking, multimodal
Anthropic nucleusiq-anthropic 0.2.2 🟢 Stable — Phase B + nested structured-output fix (additionalProperties: false on every object)
Groq nucleusiq-groq 0.1.1 🟢 Stable — Chat Completions, local tools, streaming, server-tool emission stub
Ollama nucleusiq-ollama 0.2.2 🟢 Stable — native /api/chat + multi-turn tool arguments as a mapping

nucleusiq-openai-compatible requires nucleusiq>=0.7.13. Every other first-party provider floors on nucleusiq>=0.7.12.

Cross-cutting concepts (new in v0.7.12)

Pair with the Anthropic provider guide

These three pages explain features that are exposed through nucleusiq-anthropic 0.2.0 but the observability they surface (executed_by="provider", cache_*_tokens, stop_reason, request_id) applies uniformly across every provider in v0.7.12.

  • Native server tools — web_search / code_execution / google_search / file_search — local vs provider execution.
  • Prompt caching — Anthropic cache_system / cache_tools and reading cache_read_input_tokens.
  • Extended thinking — Claude thinking="low"|"medium"|"high"|"max" and its hard constraints.