NPC-Worldwide/npcsh

What it solves

npcsh addresses the friction of switching between a traditional command-line interface (CLI) and AI assistants. It provides a single environment where users can execute standard bash commands and interact with LLM-powered agents using natural language without leaving the shell.

How it works

It operates as a composable multi-agent shell that interprets both bash and natural language. The system uses a Python backend (npcpy) to drive the inference loop and supports a wide range of model providers via LiteLLM (including Ollama, OpenAI, and Anthropic). Users can define custom agents using YAML (.npc files), Markdown (agents.md), or dedicated files in an agents/ directory. It also utilizes "Jinxes" (Jinja Execution templates) to allow agents to invoke specific tools, prompts, and skills.

Who it’s for

Developers and power users who want to integrate AI agents directly into their terminal workflow for tasks like system administration, code refactoring, and multi-step technical workflows.

Highlights

• Unified Interface: Seamlessly mix bash commands and natural language queries in one shell.
• Flexible Agent Definitions: Define agents using YAML or Markdown with customizable models and providers.
• Composable Tooling: Use Jinja Execution templates (Jinxes) for agent tool-use and skill implementation.
• Broad Model Support: Compatible with any provider supported by LiteLLM, including local options like Ollama and LM Studio.