Home

Agent Factory (af) runs many AI coding agents — Claude Code, Codex, Aider, Gemini, Amp, opencode, and Devin — and gives each one its own git branch and git worktree, so parallel agents never share a checkout and every result comes back as a branch you review like any other. You drive the whole fleet from a browser, a terminal UI, or a JSON CLI, and all three read the same state. That state belongs to a background daemon, which keeps sessions running when you close the window, schedules recurring work, and serves the web client.

Install¶

Prerequisites: tmux, git, and at least one agent CLI on your PATH, on Linux, macOS, or WSL2.

curl -fsSL https://raw.githubusercontent.com/sachiniyer/agent-factory/master/install.sh | sh

That puts af in ~/.local/bin (override with AF_INSTALL_DIR, pin with --version <tag>). Building from source needs Go 1.25+ and ./dev-install.sh.

First five minutes¶

cd your-project     # any git repo
af doctor --setup   # check tmux, git, agent CLIs, storage, daemon health
af                  # open the TUI

Press n, name the task, and describe it — the agent starts in a fresh worktree. Watch it in the Agent tab, press ↵ to type to it in place, ctrl+] to step back out, and a to archive it when you are done. Then open http://localhost:8443: the same sessions, live, in a browser, with no token and no login screen.

Getting started

The mental model in five terms¶

Session

One agent working on one task, with its own tab strip and an explicit lifecycle: archive to finish, restore to bring back, kill to discard.

Sessions and worktrees
Worktree

The isolated checkout and branch each session gets, so agents never collide and the result is ordinary git work you can diff, push, and merge.

Sessions and worktrees
Daemon

The background process that owns all state, keeps sessions alive across restarts, and serves the web client. Every surface is a thin client over it.

The daemon
Task

A prompt the daemon delivers on its own — on a cron schedule, or on each line a watch command prints — into a fresh session or an existing one.

Tasks
Account

A named credential home for an agent, so a session runs as a specific Claude, Codex, or Gemini login instead of your ambient one.

Accounts and usage limits
All five, expanded

One page with the ideas the rest of these docs assume, and how a prompt becomes a branch you can merge.

Concepts

Three surfaces, one daemon¶

Web client

Sessions, real terminals, tabs, tasks, and config in a browser — bundled into the daemon and on by default at http://localhost:8443.

The web client
TUI

af in a git repo: a dense rail of every session, in-place interaction, and full-screen attach without leaving the terminal.

The TUI
CLI

af sessions and af tasks emit JSON, so scripts — and other agents — can create, prompt, watch, and archive sessions.

The CLI

They cannot disagree, because none of them owns state:

flowchart TB
    web["web client<br/>browser · real terminals · tabs · tasks"]
    tui["af TUI<br/>rail · Agent tab · attach · panes"]
    cli["CLI · HTTP API<br/>JSON commands · local Unix socket"]
    daemon["daemon<br/>single writer of all state<br/>sessions · tasks · usage limits"]
    sessions["sessions<br/>one git worktree · branch per agent<br/>local, docker, ssh, or your own backend"]

    web -->|"read projection · RPC"| daemon
    tui -->|"read projection · RPC"| daemon
    cli -->|"RPC"| daemon
    daemon -->|"owns"| sessions

Beyond your laptop¶

Sessions can run in a container, on another machine over SSH, or on infrastructure you provision yourself; the daemon itself can be reached from a second machine over SSH or an authenticated port.

Backends and remote access

Where to go next¶

Getting started — install, first session, first browser tab.
Concepts — the five terms, expanded.
Troubleshooting — af doctor and what to file.
Why Agent Factory · Use cases · Comparison — the longer argument, and where this sits next to the alternatives.

Run many coding agents without losing control of your repo.