dmux — Parallel Agents with tmux and Worktrees

Manage multiple AI coding agents in isolated git worktrees. Branch, develop, and merge — all in parallel.

What Is It

dmux is a dev agent multiplexer that runs multiple AI coding agents (Claude Code, Codex, OpenCode) in parallel using tmux and git worktrees. Each agent gets its own isolated working copy of your repo, so they can work simultaneously without conflicts.

"Press n to create a new pane, type a prompt, pick an agent, and dmux handles the rest — worktree, branch, and agent launch."

From the team behind FormKit, Tempo, AutoAnimate, and Drag and Drop.

What Problem Does It Solve

Running multiple AI agents on different tasks in the same repo causes conflicts — you can't easily test different approaches in parallel or compare agent outputs. dmux solves this by:

Isolating each agent in a git worktree — full working copy, no conflicts
Orchestrating the entire lifecycle — create worktree → launch agent → merge back → cleanup
Enabling A/B testing — run two agents on the same prompt side-by-side and pick the winner

Key Features

Worktree Isolation

Each tmux pane gets its own git worktree and branch. Agents work in complete isolation. When done, press m to merge back into main.

Multi-Agent Support

Claude Code — Anthropic's agentic coding tool
Codex — OpenAI's coding agent
OpenCode — open-source coding agent

Automatically detects installed agents. If multiple are available, you can choose per-task.

A/B Agent Pairs

Launch two agents on the same prompt side-by-side (e.g., "A/B: Claude Code + OpenCode"). Both get the same prompt in separate worktrees. Compare results, merge the winner, discard the other.

AI-Powered Naming

Uses OpenRouter (gpt-4o-mini, grok-4-fast) to:

Generate smart branch names from prompts (e.g., fix-auth instead of dmux-1708542312)
Create conventional commit messages from diffs
Detect agent state from terminal output

Falls back to timestamps if OpenRouter key not provided.

Smart Merging

Two-phase merge to avoid conflicts on main:

Merge main → worktree (resolve conflicts in isolation)
Merge worktree → main (clean merge)
Auto-cleanup: remove worktree and branch

Auto-commits uncommitted changes before merging with AI-generated commit messages.

Lifecycle Hooks

11 lifecycle hooks for custom automation:

worktree_created — install dependencies
pre_merge — run tests before merging
post_merge — notify on completion
before_pane_create, pane_created, before_pane_close, pane_closed, before_worktree_remove, worktree_removed, run_test, run_dev

AI-assisted hook authoring: press h and an agent walks you through creating hooks.

Multi-Project Support

Attach multiple git repos to the same tmux session. Manage frontend + backend panes side-by-side. Automatically merges nested worktrees in the correct order (deepest first).

Pane Status Detection

Monitors each agent to show real-time status:

in_progress — actively working
open_prompt — waiting for user input
option_dialog — showing options

Uses lightweight LLM analysis (grok-4-fast, free) when activity stops.

Autopilot Mode

Launch agents in fully autonomous mode:

Claude Code: --dangerously-skip-permissions
Codex: --approval-mode full-auto

Keyboard Shortcuts

Key	Action
`n`	Create new pane (worktree + agent)
`t`	Create terminal pane (no agent)
`j`	Jump to pane
`m`	Open pane menu (merge, close, rename, etc.)
`x`	Close pane
`p`	Create pane in another attached project
`s`	Settings
`L`	Toggle sidebar layout mode
`h`	Create/modify hooks with AI
`q`	Quit

Quick Start

npm install -g dmux
cd /path/to/your/project
dmux

Press n, enter a prompt, pick an agent. dmux creates the worktree and launches the agent. Press m to merge when done.

Tech Stack

Node.js · tmux · Git worktrees · OpenRouter (optional)

Requirements

tmux 3.0+
Node.js 18+
Git 2.20+
At least one agent: Claude Code, Codex, or OpenCode
OpenRouter API key (optional, for AI naming)

Why It Matters

dmux represents a new primitive for agent orchestration: parallel isolated execution with lifecycle management. Instead of manually managing worktrees, branches, and agent coordination, dmux handles the entire flow. This is especially valuable for:

Agent comparison — A/B test different agents on the same task
Parallel development — run multiple tasks simultaneously without conflicts
Safe experimentation — each agent is isolated; failed attempts don't affect main

The FormKit team built this as an internal tool for running "Codex and Claude Code swarms" and open-sourced it. The approach is complementary to tools like harness-engineering-openai which focus on single-agent scale, while dmux focuses on multi-agent orchestration.

License

MIT

harness-engineering-openai — OpenAI's approach to agent-first engineering (single agent at scale)
temporal-io-ai — Workflow orchestration for agents (durable execution)
tmux — The underlying terminal multiplexer dmux is built on
alfred-portable-openclaw — Another approach to portable agent environments

Added: February 20, 2026