One Window for All Your Coding Agents and Machines
If you're juggling Claude Code on your laptop, Codex on a home server, and some other agent on a GPU box, you know the pain. You've got terminals scattered across machines, SSH keys to manage, and no single place to see what's running where. Harness is an open-source attempt to fix that by putting every agent and every machine side by side in one window.
What It Does
Harness is a desktop app (a native one, not Electron) that acts as a command center for coding agents across multiple machines. It supports Claude Code, Codex, Cursor, OpenCode, Devin, Amp, Copilot, and seven more—running each one in its own tmux pane so scrollback, colors, and keybindings work the way you expect.
The architecture is straightforward. One daemon runs per machine and manages your agents inside tmux. That daemon dials out, which means no machine has to open a port. You link each machine with a password—no SSH keys, no Tailscale, no port forwarding. Remote terminals stream text peer to peer over WebRTC rather than pixels, so you're not running a remote desktop. Everything is end-to-end encrypted: code, keys, and keystrokes are sealed on your machine, and the relay only forwards bytes it can't read.
Why It's Cool
It doesn't wrap your agents. This is a subtle but important design decision. Harness reads transcripts and uses each vendor's own hooks rather than intercepting or reimplementing agent behavior. Your CLIs stay exactly as they are, which means you're not locked into a proprietary abstraction layer that breaks every time an agent updates.
Keyboard-first, and actually remappable. ⌘O finds any session on any machine. ⇧⌘I jumps to whichever agent is waiting on you. ⌘D splits a pane. Every key is remappable through a single JSONC file that reloads on save, and chords can go up to four strokes. If you live in the terminal, this matters more than any feature list.
A worktree per harness. You can start an agent on its own branch, keeping your working copy clean. This is the kind of thing that sounds small until you've had an agent commit half-finished work to your main branch.
It's fast, and the benchmarks are honest. On an M2 Max, ⌘N, ⌘O, and ⌘T UI actions hit 11–13 ms median. Local terminal echo is 1.1 ms median. Desktop resources sit around 0.1% of one CPU core and 304 MiB, including terminal rendering and scrollback. Notably, the README points you to the raw benchmark data—workloads, slow tails, connection failures—rather than just cherry-picking the good numbers.
Close it and your agents keep working. Because agents run in tmux on the daemon, shutting the app doesn't kill your sessions. You can walk away and come back.
Local models are supported. You can run open-weight models on your own machines with Grid, Ollama, MLX-LM, and vLLM.
It's fully open source. The app, CLI, daemon, relay, and device are all in the same repository.
The README also gestures at something beyond code—CAD, circuits, robots, games, music—through what it calls "domain-specific harnesses." That's a broader ambition than most agent managers, though the README doesn't go into much detail on how it works in practice.
How to Try It
-
Download the desktop app from harness.autonomous.ai/desktop.
-
Install the daemon on each machine you want to link. The README mentions a GPU box coming online after four setup commands, so the daemon install is meant to be short.
-
Link each machine using its password—no SSH keys or port forwarding required.
-
Start an agent. It'll run in its own tmux pane, and you can split panes, jump between sessions, and remap keys as you go.
The full source—app, CLI, daemon, relay, and device—lives at github.com/autonomous-ai/openharness. If you want to dig into the performance details, the README links to benchmark docs covering both core experiences and remote transport routes.
Final Thoughts
Harness is aimed at developers who are already running multiple agents across multiple machines and have felt the friction of managing them separately. If you're running one agent on one laptop, this is probably overkill. But if your setup has grown into a small fleet—a laptop, a home server, a GPU box, each running different agents—the combination of tmux-backed sessions, password-based linking, and end-to-end encryption addresses real problems without forcing you to change how you work. The fact that the benchmarks and raw data are published alongside the claims is a good sign that the team is more interested in being useful than in being loud.
Follow @githubprojects for more developer tools and open source projects.