One window, every project. Chat is the interface; the file tree, the editor and the todo list sit right beside it.

AgentiDE (branded A://IDE) is the endpoint of the journey my case study traces: coding with AI from copy-paste to custom harness. That arc ran from pasting ChatGPT answers by hand, to GPT inside VS Code, to Google’s Antigravity, to Cursor, to Pi, a terminal agent harness I then customized heavily, layering on skills, extensions and agent templates until the environment mattered as much as the model. Herdr put many agent sessions side by side in one terminal. AgentiDE is that whole stack with a face on it: a single, compact Electron app.

The core bet is that all projects live in one interface. Open a folder as a space and it stays open: no closing one project to reach another, no pile of terminal tabs with no memory of which was which. Each space carries its own agents, panes and editor tabs. And because the agents ride Herdr, an agent isn’t jailed to its own codebase: it can read and view sibling projects, so “this function works like that over there; replicate it here” is a prompt, not a migration project.

What’s inside

The engineering is where it earns its keep. The Electron main process talks to a detached Herdr session over its unix socket, so agents survive closing the window; chat messages are parsed from each agent’s session JSONL with raw-byte framing, never scraped from terminal text; agent creation runs a six-stage boot pipeline in the UI so there’s never a blank wait. Every space gets a file tree, a Monaco editor and viewer for code and the plans agents write, and the raw terminal one keystroke away, all scoped per project so files never leak across spaces.

Mid-work “oh, and I need this too” is constant, so todos live in the IDE itself: add one while the agent is working, and a single click starts a new agent on that task in the right project.

Four points of proof

Every project onboards through four layers that together prove out a codebase before an agent touches it. A codegraph substrate for the mechanics: structured symbol, caller and blast-radius queries over how the code actually functions. A live DBML schema layer with an ERD panel for the ongoing view of the database: tables, structure, how everything reacts. Git, always used and always current, as the record of what changed. And living docs that carry the intent behind the code. Lookup routing sends mechanics questions to the graph, schema questions to the DBML, and intent to the docs.

A Pi Library surface manages every skill, extension, tool and agent template in one place, and new ones can be scaffolded by chatting with the built-in ops agent.

Why I built it

I direct agents all day, and the tools for it were a pile of terminal panes and tabs with no memory of which project was which. I wanted the workflow I actually use, productised for myself: every project in one window, agents that can see each other’s codebases, a chat-first interface, todos that turn into agents on a click, and knowledge layers that stop agents guessing how the code fits together.

It ships as a signed, notarized macOS build with a fail-closed release pipeline, a reversible installer and a 450-plus test suite. It’s my daily driver; there’s no public site, it’s just for me.