Irie

How it works

How Irie works: you give Irie tasks, Irie starts an agent for each inside tmux, and the agents report back through hooks. At the top is you. An arrow labelled tasks runs down to Irie, and one labelled who needs you runs back up. Below Irie, an arrow labelled starts runs down into a dashed box labelled tmux, which keeps them running, and one labelled hooks runs back up. Inside tmux are three Claude Code agents: one working, one that needs you, and one done. Under each agent is a folder labelled worktree: its own working copy of the project. You tasks who needs you Irie starts each agent and tracks it starts hooks tmux keeps them running Claude Code working Claude Code needs you Claude Code done worktree worktree worktree each in its own worktree
You talk to Irie. Irie starts your agents, keeps them running and listens for news from them.

You type a task

At the bottom of Irie’s list is a box for tasks. Press Tab to get there, type what you want done in plain words, and press Enter.

Irie starts a coding agent, an AI program that works on code, on the task. A row appears in the list, named after the task’s first few words.

Unless you name a project or a folder with @, each task gets a new, empty folder to work in.

Each agent runs in tmux

Irie starts every agent inside tmux, a free program that keeps programs running in the background. Close the window and your agents keep working.

Irie keeps a tmux of its own, apart from any you already use.

You look in when you want

Pick a row with the arrow keys and press Enter. That agent’s live screen moves into the big space on the right, where you can read it and type to it.

Space shows it but keeps you in the list, and Ctrl+] hops between the list and the agent.

Recording to come: pressing Enter on a row, and that agent’s screen moving into view.

Agents tell Irie how they’re doing

Agents report through hooks: small commands an agent runs at set moments, such as just before it runs a command or edits a file. From them, Irie knows who’s busy, who’s done and who’s waiting for you.

Irie adds its hooks each time it starts an agent, and never edits the agent’s settings files.

What the colours mean

The states an agent can be in, with their marks and colours, as Irie’s list shows them. Working, a plain dot. Needs you, a red dot and red name. Stalled, a red exclamation mark. Failed, a red cross. Done, a green tick. Idle, a dim ring. Starting, a dim dotted ring. Exited, a dim small dot. Paused, two yellow bars, planned. Frozen, a cyan diamond, planned. Irie working needs you stalled failed done idle starting exited paused planned frozen planned
Red means it needs you. Green means it’s done.

Most states have a mark of their own as well. Working and needs you share a dot, and red tells them apart.

In your terminal, the window where you type commands, the state colours are its own red, green, yellow and cyan, so they suit your theme.

Four colour themes for Irie itself are planned: Rosewater, Terracotta, Lilac and Evergreen. See them in light and dark. Planned

Each agent gets its own worktree

git keeps the history of a project’s code.

A worktree is a second working copy of a project. It shares the project’s history, but has its own files and its own branch, or line of changes.

Start a task with @ and a project’s name, and its agent gets a fresh worktree, so agents keep out of each other’s way. You tell Irie once, in its settings, where your projects live.

Irie copies in the files git leaves out, such as a project’s private settings.

Version 1 never deletes a worktree or a branch. Worktrees, in the docs

Profiles: which agent, which model

A profile is a named recipe for starting an agent: which program, which model (the AI itself) and which settings. Every task starts from your default profile: out of the box, Claude Code with Opus 5.5 at max effort in auto mode.

You can change the default in Irie’s settings, and pick another for one task with + and its name. Profiles, in the docs

The program is the agent’s harness: it turns a model into a coding agent.

Irie can start any agent that runs in a terminal. Only Claude Code tells Irie how it’s doing, so for other agents a row shows only whether they’re still running. Other agents, in the docs

Fuller support for more agents is planned after version 1. Planned

What stays on your computer

Irie runs on your computer and sends nothing over the internet of its own. Each agent talks to its own AI company, just as it would without Irie.

Irie keeps a log of what each agent did, such as the commands it ran and the files it read or changed, on your computer for 30 days.

Irie never reads or passes on your logins. It starts each agent under your own account.

What Irie is lists everything it keeps and sends, and how to check it yourself.

Coming next

Tidy the list with park, rename and undo. Then pause an agent at its next step, freeze one at once, mark a task as urgent, and watch the helpers an agent starts.

The roadmap has them in order.