Sign in Subscribe
news

OpenHarness

Sebastien Dubois

Sebastien Dubois

— 4 min read

Canonical version: OpenHarness.

OpenHarness (the app itself is just called "Harness") is an open-source command center for coding agents, built by Autonomous (autonomous.ai). Their tagline: "The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center."

It runs the agent CLIs you already use (Claude Code, Codex, Cursor, OpenCode, Devin, Amp, GitHub Copilot, Hermes, Grok Build and a few more) in one window, on whichever machine they live on: your laptop, a home server, a GPU box in the closet.

The repository went public on 2026-08-04 under the MIT license and passed ~1.1k GitHub stars within two months.

The problem it solves

With a few agents running, you lose track of them. Claude Code at work, Codex at home, Hermes on a cloud box, each in its own terminal, on its own machine, behind its own SSH setup. It gets hard to tell which session is where and which one is waiting on you.

OpenHarness puts all those sessions in one place and flags the ones that need you. Conductor, Superset and Paseo go after the same problem; OpenHarness does more on the multi-machine side.

How it works

In the app, a harness is one running agent session (its own conversation and working context), and a tab groups harnesses.

  • One daemon per machine runs your agents inside tmux. It dials out, so no machine opens a port
  • Linking a machine takes a password: no SSH keys, no Tailscale, no port forwarding. Install the CLI with a curl script, then run harness login, harness remote-password set and harness start
  • Traffic takes one of three paths, picked automatically: a direct WebRTC channel between machines, Cloudflare's TURN network when a firewall blocks that, or Autonomous' own relay as a fallback while WebRTC negotiates
  • Everything is end-to-end encrypted (ChaCha20-Poly1305, X25519 session keys, pinned Ed25519 identities). The relay forwards ciphertext and holds no keys
  • Remote terminals stream text peer to peer; there's no remote desktop involved
  • Your agents aren't wrapped. Harness reads their transcripts and uses each vendor's own hooks
  • Each harness gets its own worktree and branch, so your working copy stays clean
  • Local models run through Grid (Autonomous' own model manager), Ollama, MLX-LM and vLLM
  • The same install gives you hn, a TUI that keeps tmux's keys and your ~/.tmux.conf and shows every harness on every machine

An experimental "Tab collaboration" setting lets agents in the same tab consult each other. It's off by default, and I'd leave it off until I know what it does to the token bill.

Native app, keyboard driven

The app is built with Flutter, not Electron. Closing the window doesn't stop the agents, since they live in tmux under the daemon.

Published idle numbers on an M2 Max: ~0.1% of one CPU core and 304 MiB with 16 terminals open, and 11-13 ms median to open UI views. ⌘P finds any harness on any machine, ⇧⌘I jumps to the one waiting on you, ⌘D splits. Every key is remappable in a single JSONC file.

Downloads exist for macOS and Linux. No Windows build at the time of writing.

Domain-specific harnesses

A domain-specific harness (DSH) turns a coding agent into a specialist for one craft. It bundles instructions and skills, a pinned toolchain, a project template, checks and a live viewer: you chat on one side, and the part, circuit board or game takes shape on the other. A DSH is just a folder with a harness.json file, so adding a craft never touches the app. The other agent dashboards I know of stop at code.

The Store lists 50 of them at the time of writing, e.g.:

  • Design: Blender, FreeCAD, OpenSCAD, Excalidraw, text-to-cad
  • Engineering: KiCad (PCBs), CircuitJS, Yosys, Orca Slicer, Home Assistant
  • Simulation and robotics: MuJoCo, drone and CFD agents
  • Games: Godot (via Godogen), Phaser, voxel worlds
  • Music: Strudel, Ableton, JUCE
  • Science and data: RDKit (chemistry), marimo, data studio
  • Documents: Typst, Marp

The premise is that code is the common medium. Geometry scripts make parts, netlists make boards, animation code makes film. If a coding agent can write the script AND you can see the result live, it can work in any discipline that has a scriptable tool. It's also a good argument that the harness matters as much as the model (see Harness Engineering): same model, different toolchain and viewer, different craft.

The Harness device

Optional, and fully open hardware: a round, always-on AMOLED display (ESP32-S3) that sits next to your keyboard and shows what each agent is doing, which one finished and which one is waiting on you. You can answer an agent's question on the screen, or tap and speak a new task. It talks to the host's daemon over USB (no Wi-Fi setup, no account on the device).

You can buy one from Autonomous or build your own: the repo ships the firmware, the PCB files (schematic, Gerbers, BOM) and STEP/STL files for the enclosure.

Compared with other tools

  • Claude Code, Codex CLI and the other single harnesses are the agents. OpenHarness sits on top of them
  • Conductor and Superset run parallel agents in worktrees on one Mac. OpenHarness adds more machines, the encrypted remote layer and the non-code crafts
  • Paseo is the closest (a daemon plus clients, provider-agnostic). Paseo leans on mobile and voice, OpenHarness on multiple machines, a native desktop app and a physical device
  • OpenHarness covers what a tmux + Tailscale setup does. If yours already works, you'd gain the "who's waiting on me?" view and the DSH viewers

Caveats

  • Young project: public for about two months at the time of writing, and moving fast
  • harness login means an account with Autonomous, and their relay sits in the path whenever direct WebRTC fails. Per the encryption design it can't read anything, but you still depend on their infrastructure for discovery and fallback
  • Autonomous sells hardware (the device, among other products), so the open-source app is partly a funnel
  • DSH quality will vary a lot. Upstream tools keep their own licenses, and a wrapper around KiCad or MuJoCo is only as good as the instructions and checks it ships

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe