I’m building Agentic Control Freak (ACF): a local control plane for Codex CLI, Claude Code, Antigravity CLI, and Ollama. With ACF, you can build local projects ranging from Next.js apps to training ML models all via your browser on your local machine.
The agents still do the coding. ACF owns the outer loop:
“task -> plan -> approve -> execute -> verify -> preview -> handoff”
It keeps durable state outside any one agent session: plans, tasks, filesystem diffs, checkpoints, undos, forks, verification runs, previews, memory, skills, and handoff briefs. That means you can switch providers mid-project without starting from zero.
It also does not trust the agent’s final message. The filesystem diff is truth; verification and live preview are owned by the control plane.
The orchestration, workspaces, previews, and state are local: localhost-only, workspace-confined under .workspace/<project>, and not meant for hosting. Provider CLIs may still call their model services. Optional Telegram control lets you approve or steer runs from your phone through a token-gated local worker.
Repo: https://github.com/Antibody/Agentic-Control-Freak
Video intro: https://youtu.be/1fRH-XQrgkY
ML model design and training video: https://youtu.be/KdiiU4RIfFU
P.S. I’d love feedback, especially on whether the app works on Mac or Linux (I tested only on Win11).
I don’t know if I love it, but it’s a lot of fun. More fun than reviewing the code and fixing all the bullshit in the output. That’s literally the worst part of software development.
I usually also review my own code, and when I wrote it myself, it’s orders of magnitude faster because I recognize what I’ve written, too, so I’m minimizing the boring part as well. And being more efficient.
I just don’t get it.