# agentmaker

A Python framework for building LLM agents and multi-agent systems, with tools, memory, retrieval, guardrails, human-in-the-loop review, and a built-in run debugger.

agentmaker is a general-purpose Python framework for building LLM agents and multi-agent systems. It supports chat and tool-use loops, plan-and-execute and reflection workflows, declarative agent specs, and multi-agent orchestration, with native support for OpenAI, Anthropic, Gemini, DeepSeek, and other providers including local models.

It includes hybrid retrieval and a full RAG pipeline with source citations, long-term memory that extracts and updates facts, input and output guardrails, an approval gate for high-risk tools, and multi-tenant scoping. Context engineering assembles each prompt under a token budget with history compaction.

Every backend, including embeddings, vector store, reranker, and session stores, sits behind a swappable interface, and a built-in test double lets agents run in CI with no API key. A built-in debugger traces runs to pinpoint a failed step's root cause.

Website: https://agentmaker.xinhuang.me/