Autonomous  ·  Rust  ·  MIT

Give your AI agent
a whole team.

superharness wraps tmux orchestration around Claude Code, OpenCode, or Codex — spawning parallel workers, surfacing panes that need your attention, and keeping work flowing while you step away.

your-project — superharness
$ superharness AI agent starting — orchestrator mode You: build a REST API with JWT auth [orchestrator] Breaking into 3 parallel workstreams... [worker-1] spawned → implementing JWT middleware [worker-2] spawned → building user CRUD endpoints [worker-3] spawned → creating database schemas & migrations [heartbeat] worker-2 waiting — surfacing pane [worker-2] asks: allow file write to src/routes/users.ts? [orchestrator] approved ✓ (hiding pane) [worker-1] done · [worker-3] done · [worker-2] 80%... [orchestrator] all workers finished — branches merged, cleanup done

How it works

Install once. Works with Claude Code, OpenCode, or Codex. The rest is automatic.

1

Install superharness and run it in your project

One command: curl -fsSL https://superharness.dev/install.sh | sh. Requires tmux and your AI coding agent (Claude Code, OpenCode, or Codex). Navigate to your project, run superharness, and you're set.

2

Tmux management: surface, approve, relay

superharness actively manages your tmux panes — automatically surfacing workers that need attention, relaying questions to you, approving safe operations on your behalf, and hiding idle panes out of the way. A heartbeat loop detects stuck workers before they stall.

3

Away mode + task tracking

Step away and superharness keeps running. It auto-approves safe operations, queues uncertain decisions, and debriefs you when you return. Tasks are tracked from breakdown through spawning to completion — you always know what's in progress, blocked, or done.

What you gain

Your AI agent, now running a whole team — with you always in control.

Parallel workers

The AI spawns as many worker agents as the task needs — each tackling an independent workstream simultaneously. What takes one agent 4 hours takes a team of 4 about an hour.

🖥️

Smart tmux management

superharness surfaces worker panes that need attention, hides idle ones, and keeps the workspace clean. You always see what matters — and nothing you don't.

💓

Heartbeat loop

A continuous heartbeat detects stuck or waiting workers before they stall your session. superharness wakes the orchestrator the moment a worker needs a decision, approval, or recovery.

🌙

Away mode

Step away and superharness keeps working. It auto-approves safe operations, queues uncertain decisions, and gives you a full debrief when you return — no context lost.

📋

Task tracking

Work is broken into tasks and tracked from spawn to done. See what's running, blocked, or finished at a glance — F5 shows the live task board.

🔒

Isolated workspaces

Every worker gets its own git worktree. No file conflicts. Workers commit their own branches; the orchestrator merges results and cleans up automatically.

Single agent vs. superharness team

Your AI agent is great. Now give it a team.

solo agent
  • One task at a time, sequential
  • Slow on complex multi-part work
  • Every permission prompt reaches you
  • Stalls when waiting — no recovery
  • One context window for everything
  • Nothing happens when you step away
superharness
  • Orchestrator spawns workers in parallel
  • Each worker gets an isolated worktree
  • Safe operations auto-approved, rest surfaces to you
  • Heartbeat detects and recovers stuck workers
  • Each worker gets a fresh focused context
  • Away mode keeps work flowing while you're gone

Install

Requires tmux and an AI coding agent (opencode, claude, or codex).

Quick install
$ curl -fsSL https://superharness.dev/install.sh | sh
Ubuntu / Debian (PPA)
$sudo add-apt-repository ppa:borodutch/superharness
$sudo apt update
$sudo apt install superharness
Arch Linux (AUR)
$ yay -S superharness
macOS / Linux (Homebrew)
$ brew install backmeupplz/superharness/superharness
Cargo
$ cargo install superharness
then, in any project
$cd your-project
$superharness

Going somewhere?

Leave the AI running. It handles things while you're gone.

Away mode

The orchestrating AI can enter an away state when you step out. In away mode it keeps workers running and approves safe operations, but queues any real decision for your return — architecture choices, destructive operations, anything it isn't sure about. Come back to a full debrief of what happened and a short list of things that need your call.