https://arxiv.org/abs/2506.10077
followup paper coming soon which further demonstrates these contextuality results for a suite of models. there is no way to fundamentally impose on the training data or processing effective guardrails that can transcend this reality.
from article: "NPC Shell transforms your terminal into an intelligent workspace that extends the traditional bash environment with AI capabilities. Inspired by biological intelligence, it provides a unified command-line interface for seamless interaction with multiple LLM providers and specialized agents -- enabling multi-modal collaboration, code execution, and automation of complex tasks."
npcsh is a command-line toolkit that provides simple and practical ways to use AI tools wherever you work.
npcsh lets you easily tweak agent personas, tools, and context. It comes with several agents that you can work with:
1. guac: a python coding agent that executes code that you can inspect and work with. Likewise, guac can build on the code you write within the session and works by way of managing a python package to help you go from experimentation to automation more quickly
2. corca: a tool-calling agent that can hook up to an mcp server. By default, corca uses the mcp server we ship with npcsh, but we've used it with other mcp servers as well.
3. alicanto: a deep research agent that runs experiments and then writes a paper in a latex document, and, if equipped with something like a semantic scholar api key, alicanto can even find citations and incorporate them into the work
Additionally, users can take advantage of helpful macros like '/ots' for screenshot -> LLM, '/sample' to get a one-off answer that has no prior context, '/vixynt' to generate or edit images.
with npcsh, we are building not just helpful agents, but the toolkit with which you as a user can continually improve your own agent teams.
https://arxiv.org/abs/2506.10077 followup paper coming soon which further demonstrates these contextuality results for a suite of models. there is no way to fundamentally impose on the training data or processing effective guardrails that can transcend this reality.
reply