I’ve been building resh, an open-source, AI-native resource shell focused on making infrastructure automation structured, typed, and machine-friendly instead of brittle text pipelines.

resh treats system resources as typed handles instead of raw commands:

file:// – files & storage

proc:// – processes

svc:// – services (systemd + OpenRC)

mq:// – local message queues

net:// – networking diagnostics

http:// – real HTTP access (not a wrapper scheme)

log:// – structured logs

secret:// – typed secrets

snapshot:// – host snapshots

config:// – config state

tag:// – metadata tagging

Each handle exposes verbs with deterministic JSON output designed for both humans and LLMs.

What’s new in v0.7:

systemd + OpenRC service control (start/stop/restart/reload/enable/disable/mask/is-enabled/wait)

typed HTTP verbs (get/head/post/put/patch/delete/json/options)

AI-friendly body typing (json/text/bytes w/ base64)

atomic file operations (write/append/rename)

secret backends (local/env with rotation)

snapshot diff + restore

consistent --json-pretty output support

The goal is to make automation verifiable, introspectable, and safe for AI agents without brittle parsing.

Project is fully open source and early but usable. I’d love feedback on:

the handle/verb model

missing primitives

AI use cases that would benefit most from this style

Repo: https://github.com/millertechnologygroup/resh

Thanks for checking it out — happy to answer any technical questions.