luyi14-bits/jig

Jig is a Python framework for building and coordinating multiple AI agents that work together on a task. It is written specifically with DeepSeek's models in mind, though it also supports OpenAI and can be extended to other providers. The project is in an early alpha stage (version 0.5.0), so the README is upfront that it is not yet meant for production use. The core idea is a safety layer called ToolGuard, which checks every tool call an agent wants to make before it actually runs, rather than after. Jig also includes a four-layer memory system, a graph-based way to arrange how agents pass work to each other, and support for streaming responses back to the user as they are generated. A cost-aware router picks cheaper or faster models for simpler steps to help manage spending. Out of the box, Jig comes with twelve preset agents covering roles like project management, coding, code review, test-driven development, security review, and DevOps, each defined through a markdown file called SKILL.md. You can use these as they are, edit them, or write your own agent by creating a new SKILL.md file that describes the agent's role, its step-by-step workflow, and any rules it should follow. Once added to the skills folder, Jig picks the new agent up automatically. Getting started involves installing the package with pip, setting an API key as an environment variable, and calling a Python function with the task you want done. The dispatcher then routes that request through the appropriate agents. The project ships with documentation covering a technical whitepaper, a user guide for the command line and FastAPI server, and a comparison against other agent frameworks like LangGraph and CrewAI. It includes 117 automated tests. The license is MIT.