Bring multi-agent workflows and context engineering into your daily shell.
NPC Shell (npcsh) is a command-line first toolkit that enables composing multi-agent workflows, Jinja templates, and context files directly within the terminal. It supports both local model stacks and cloud APIs, making it suitable for rapid prototyping, team collaboration, and edge deployments.
Detailed Introduction
NPC Shell organizes NPCs (agents), Jinxs (macros), and team context (.ctx/.npc) as composable data structures. Users can call, switch, and combine different agents from the shell. The project provides a rich CLI, examples, and documentation to run multi-step automated tasks locally or in CI. See the official docs for more.
Main Features
- Shell-first interaction: converse with agents via natural language or CLI commands.
- Multi-agent orchestration: build and switch between role-based NPC teams.
- Local and cloud compatible: supports local inference stacks and common cloud providers.
- Extensible data layer: use .npc/.jinx/.ctx YAML files to manage context and templates.
Use Cases
Useful for developers who want to iterate on prompt engineering in-terminal, automate documentation edits, build agent pipelines, or deploy lightweight agent services in constrained networks. Project context and components can be versioned for team sharing and reproducibility.
Technical Features
Implemented in Python, npcsh offers a modular CLI and plugin hooks. It integrates LiteLLM-style providers to support multiple model backends. Install options include a lightweight API mode or a full local stack (diffusers/transformers) for on-prem inference. The repository is MIT-licensed and accepts contributions via GitHub.