I had a setup that worked perfectly for me,
3 screens (Actually 2 screen + Macbook Pro)
one with IDE (WebStorm, Vscode and from time to time just sublime).
Second with Terminal (started with iTerm2 and moved to Warp with Oh My Zsh and bunch of plugins)
and last with Browser (Web or DB).
Add to that the fact that I’ve always been a sucker for good GUIs,
so GitKraken for Git and Lens for Kubernetes completed most of my stack. It worked extremely well for years.
But when I started writing less code myself and letting agents write more of it for me, I realized this setup no longer worked.
Mixing AI into this flow (chatgpt, gemini & claude) started a butterfly effect.
My appetite has increased, I wanted to scale myself, run in parallel on many features and repositories but I was missing a way that felt natural and comfortable for me.
The moment I started jumping between multiple sessions/projects (Hey Tmux/Cmux), I got lost. More than once, I sent unrelated prompts to the wrong project or the wrong session. Creating worktrees and tracking them became messy, running multiple local envs and finding which of the port is running which branch got me dizzy.
It became obvious: this needed a proper methodology and a control plane.
Now Spec-Driven Development gave a strong theoretical answer on how you can have a system to plan and let agents execute, plan many specs let agents run and… 🎩🐇.
On paper, it felt exactly right for this new world. But when I tried the tools that exist today, they felt too clunky and too slow for me.
I had already seen how quickly I could get to clear definitions through chat or plan mode, and with the right skills, tests and loops in place, the results were genuinely impressive.
I ended up having many conversations about this with colleagues, how to take what we already know works in structured software development and make it comfortable first and foremost for us.
That’s exactly how Shep was born. A shepherd for a flock of agents.
The idea is simple: if agents are becoming part of how we build software, we need better ways to organize them, guide them, and keep context from falling apart.
According to Steve Yegge, people building tools like this are already in a later stage of the AI-assisted development journey. Maybe these tools are only for people at that stage. Maybe not.
But for me, this came from a very practical pain: I didn’t want more agent power. I wanted a better way to work with it.
Shep runs locally, doesn’t require a complicated setup, and tries to fit naturally into how different developers already like to work. It supports Claude and Cursor, Gemini & Codex, works on Mac, Linux, and Windows, open source under MIT no token, api key or register needed, no heavy platform lock-in.
There’s a lot more in the repo for anyone who wants to go deeper, but if you just want to try it:
npx @shepai/cli
Shep AI
One command. Full lifecycle. Merged PR.
Describe a feature in plain English — Shep researches, plans, codes, tests, and opens a PR. You approve when you want to, or let it run hands-free.
Features · Quick Start · CLI Reference · Architecture · Contributing
Quick Start
# Try it instantly — no install needed
npx @shepai/cli
# Or install globally
npm i -g @shepai/cli
shep
# Browser opens at http://localhost:4050 — you're in
Features
shep feat new "add stripe payments" --allow-all --push --pr
# ↳ PRD → research → plan → code → tests → PR → CI watch — done.
- Full lifecycle in one shot — From idea to merged PR: requirements, technical research, implementation plan, code with tests, PR creation, and CI fix loop
-
Approve or go hands-free — Three review gates (PRD, Plan, Merge) you can enable, disable, or skip entirely with
--allow-all - …
Would genuinely love feedback especially honest criticism from people trying to make AI-assisted development actually usable in day-to-day work.
Top comments (0)