One window, many agents. Meet onethread

Problem

A tab per agent. A tab per task. A tab per repo. That’s the nightmare most of us live in now that coding agents are everywhere.

Each agent waits in its own tab. You walk from one to the next, rebuild the context in your head, answer a question, and move on, until you can’t remember which tab asked what.

How many agent tabs do you have open right now?

How kings worked

As humans, we’re used to working on one thing at a time. That’s why kings didn’t walk from task to task. Tasks came to the king, and he delegated.

I think that’s still the most natural way to coordinate. So I built it.

Meet onethread

onethread is a new fruit in my garden. You talk to one agent. Many coding agents work behind it. Version 1.0.0, the first stable release, is out now.

Run ot, pick Claude Code or Codex, and talk in that window the way you always do. That window is the orchestrator. It keeps a queue of tasks across your repos and runs each task’s own agent session in the background: up to 10 at once, each in its own git worktree. When something is ready, it tells you at a natural pause, never mid-thought.

A right hand who remembers

The orchestrator holds the context, so you don’t have to. It understands you across all those tasks at once. Say “PR 17” or “the PR that adds shortcuts”, and it knows which one you mean.

When you do want to step into a task yourself, ask for it by name. A new iTerm2 tab opens inside that task’s real session. If its agent was working, it stops and greets you with what it was doing.

Your kingdom, your rules

You choose how your agents behave and how they address you. Run ot me king, or Batman, or anyone at all, and the orchestrator builds your people from that world.

I’m King Igor. Leonardo da Vinci is my right hand, and my court includes Michelangelo, Pythagoras, Newton, and Einstein, and also Putin and Trump. You can even have Jesus as a servant in your kingdom. Your imagination is the limit, and work becomes a game.

Real people are played as a light parody, and the game stops where it should: code, ids, errors, and confirmations stay plain. ot me plain turns it off.

The king still signs

Delegating doesn’t mean handing over the seal. Merge, push, cancel, and accept ask you in the terminal or a macOS dialog, and the model can’t answer for you.

It’s all local: Python 3 standard library, one SQLite file, and the claude or codex CLI you already use. No server, no account, no telemetry.

Limits

onethread needs macOS, Python 3.9 or newer, git, and at least one of Claude Code or Codex CLI, installed and logged in.

iTerm2 is recommended for task tabs and notifications. Codex can’t be woken, so it tells you what happened on your next message.

Workers can run a repository’s test and build scripts, so only use repositories you trust. Some real-tool checks are still outstanding; docs/STATUS.md in the repository lists them.

Try it

See it first: the demo runs the full scenario with simulated agents and spends no quota.

git clone https://github.com/RogyMD/onethread.git
cd onethread
python3 demo/run_demo.py

Then install it and start talking:

./ot install
ot

./ot install asks before it links ot into ~/.local/bin. onethread is MIT-licensed.

Closing

Stop managing tabs. Manage the work.

Meet onethread: https://aigarden.uk/onethread

Read the v1.0.0 release: https://github.com/RogyMD/onethread/releases/tag/v1.0.0

Everything else, including how to remove it, is in the README: https://github.com/RogyMD/onethread#readme

Tell me who you’d crown yourself, and who’d sit at your right hand.

Best wishes 🌊
from AI Garden