autonomous-harness

How we running Claude across multiple machines without the window-hopping When you start scaling Claude workflows beyond a single machine, like running heavy coding tasks or Claude CLI sessions on a main desktop while keeping a secondary rig or an old laptop humming away in the corner, monitoring things gets messy fast. You end up constantly jumping between SSH/VNC windows or physically turning around just to check what each machine is doing. I work at Autonomous. During our recent Product Hunt launch, a dev hit the nail on the head regarding the other half of this friction: "If an agent stops to ask something while I'm away, it just sits there until I get back". The dead time is a massive bottleneck when you're trying to run agents asynchronously. That exact combination of issues is why our team built Harness. It’s an open-source app designed around two main goals: - Unified Multi-Machine View: Pulling agents running across different computers into a single window so you can see what they're all up to without jumping between machines. - Mobile Routing: If Claude hits a prompt and needs input while you're away from your desk, the question routes straight to your phone instead of idling blindly in an unattended terminal. Under the hood, processes run natively on your machine rather than inside a closed sandbox, meaning closing the app doesn't interrupt your tasks, and your agents keep using their own local accounts and native hooks. It’s MIT licensed and works on Mac and Linux if anyone is dealing with a similar multi-rig setup. Here is a closer up on how we made it: autonomous-ai/openharness What's your current workaround when an agent stalls for input while you're away from the keyboard? Do you just stay glued to the terminal until it finishes or have you built custom alerts/scripts to let you walk away?

10 projets GitHub qui ont l'air faux (et pourtant ils existent) Un clavier dont les touches viennent jusqu'à ta main, un robot-crabe qui danse, Hollow Knight sur une calculatrice NumWorks… Voici les 10 trouvailles que vous avez le plus regardées cette semaine sur GitTendance. Pour chacune : comment ça marche, ce que ça coûte, et le hic. 👉 Les liens 10. Compass : AgriciDaniel/compass 9. Strata : Niko1221/Strata 8. DroidDeck : Droid-Deck/DroidDeck (pas de site officiel : télécharge uniquement sur GitHub) 7. OpenHarness : autonomous-ai/openharness 6. Map3D : cartesiancs/map3d 5. SkyCraft : chasmlol/SkyCraft 4. NumPlay : Mason363/NumPlay 3. Xbox Nano Controller : news.xbox.com 2. Jumper : KingKongRobotics/jumper 1. Gboard Kurukuru : github.com/google/mozc-devices/tree/main/mozc-con… Juste en dehors du top : DigUp : ARahim3/DigUp Filee : KnifeLemon/Filee Summer Cycle : StarKnightt/summer-cycle Et toi, lequel tu installerais en premier ? Dis-le en commentaire 👇 Chapitres 0:00 Intro 0:23 10. Compass 1:16 9. Strata 2:14 8. DroidDeck 3:06 7. OpenHarness 3:57 6. Map3D 4:43 5. SkyCraft 5:30 4. NumPlay 6:24 3. Xbox Nano Controller 7:03 2. Jumper 7:53 Juste en dehors du top 8:43 1. Gboard Kurukuru 9:27 Le classement complet Images et vidéos : Google Japan, Xbox, Autonomous, chasm (@chasmmm), @circuits24x7 sur X, Daniel Agrici, Google Gemma, KingKong Robotics, Niko1221, Droid-Deck, Mason363, Filee, Summer Cycle, © les contributeurs d'OpenStreetMap. Chaque média est crédité à l'écran. #github #opensource #ia #gaming #tech #top10

We built Harness, a device to put all of your agents in one place I want to share what we’ve been working on and why. As we started using more coding agents, they ended up spread across different terminals and machines. I might have Claude Code open on my laptop, Codex on another computer, and a longer task running on a server. The agents could do useful work in parallel, but keeping track of them meant checking windows and remembering where each task was running. Harness is how we try to make that less fiddly. It gives each agent its own session and worktree, and puts sessions from different machines in one place. The agents still run where they normally run, in their usual CLIs and terminals. Harness is there to start them, follow their work, and make it easier to respond when one needs input. The desktop app is the main workspace. There’s also a terminal interface and a mobile app, so you can check on work or reply without opening every session. We made a small desk device for the same reason. It shows what’s running and what has finished, and lets you give an agent a task by voice. The idea is to make status visible without adding another full chat window to your desk. We’ve also been building beyond coding. A “harness” can give an agent the tools and viewer for a particular kind of work, like Blender, circuits, robotics simulations, or games. The agent handles the reasoning; the harness gives it a more useful workspace and gives you a way to see what it’s making. A lot of the work has been in the plumbing: keeping sessions alive when the app closes, connecting machines without asking people to set up port forwarding, and making sure each agent’s work stays separate. Open Source: The project is open source, including the app, CLI, daemon, relay, device firmware, and hardware design files.

We built Harness to put all of your AI agents in one place I want to share what we’ve been working on and why. As we started using more coding agents, they ended up spread across different terminals and machines. I might have Claude Code open on my laptop, Codex on another computer, and a longer task running on a server. The agents could do useful work in parallel, but keeping track of them meant checking windows and remembering where each task was running. Harness is our attempt to make that less fiddly. It gives each agent its own session and worktree, and puts sessions from different machines in one place. The agents still run where they normally run, in their usual CLIs and terminals. Harness is there to start them, follow their work, and make it easier to respond when one needs input. The desktop app is the main workspace. There’s also a terminal interface and a mobile app, so you can check on work or reply without opening every session. We made a small desk device for the same reason. It shows what’s running and what has finished, and lets you give an agent a task by voice. The idea is to make status visible without adding another full chat window to your desk. We’ve also been building beyond coding. A “harness” can give an agent the tools and viewer for a particular kind of work, like Blender, circuits, robotics simulations, or games. The agent handles the reasoning; the harness gives it a more useful workspace and gives you a way to see what it’s making. A lot of the work has been in the plumbing: keeping sessions alive when the app closes, connecting machines without asking people to set up port forwarding, and making sure each agent’s work stays separate. The project is open source, including the app, CLI, daemon, relay, device firmware, and hardware design files. The product was launch on Product Hunt and rank #2 on that day. If you try it, I’d be interested to hear what feels awkward or what you’d want it to handle next. Open Source: everything is open, some of our followers have already made one for their own

OpenHarness: An app to harness all your agent, update after 1 month Disclosure: I work on OpenHarness. MIT, everything still in one repo. I posted here about a month ago running agents on three machines from one window. I ended that post saying Windows was on the list and I wouldn’t put a date on it. Update on that, and on a few things I’d rather tell you than have you find in the commit log. Windows is merged. PR #1470 went into main: terminal host, CLI, desktop app, dial and installer. One x64 installer that picks the build for the PC it runs on. CI spins up a Windows VM, installs it, checks it, uninstalls it, then checks again that nothing is left behind. The last fixes before merge were things like a standard user over SSH hitting WMI refusing a network logon, so the uninstaller couldn’t finish its last step and still reported success. That kind of bug is the reason I wouldn’t give you a date a month ago. Two changes you should know about before you decide whether to keep using it: Sign-in is now required on Desktop and TUI. It wasn’t before. Everything else from my last post still holds, the daemon still dials out, you still open no ports, the relay still can’t read what it passes. But an account is now part of the setup, and if that’s a dealbreaker for you then it’s a dealbreaker and you should know now rather than after install. Desktop setup counts are collected by default, with the opt-out in Settings. Same deal: I’d rather you hear it here. What else landed in the last month: - Multiple machines from the CLI. --machine names another machine, and you can read and wait on its harnesses without switching context. - Super Tab. Terminals group together, with a team roster and machines other people contribute. Shared conversations. Every delegation gets routed, logged, and can be reassigned by a person afterwards. Shared tabs are discoverable across computers. - Cloud agents as harnesses. Fo, Devin Cloud and Cursor Cloud sit next to your local ones, and the CLI can start, read and answer them. If you’re here because you want everything local, you can ignore this entirely, nothing changed about the local path. - Guest machine sharing. A guest gets their own OS account on the host and sees only the folders you pick. - MCP server. Added harness_search, harness_machines and harness_whoami. harness_start runs any engine now. Namespace moved to ai.autonomous.harness. - Desktop performance. No more 30 second rebuilds, hash updates and notifications moved off the UI thread, and slow frames are now counted rather than guessed at. Repo: autonomous-ai/openharness  

@dreamworks2050 yeah fully open source all the 3d stl/step, schematics, firmware, etc autonomous-ai/openharness

@Teknium Hermes is Day-0 support on all of our devices (Harness, Lamp, Intern, etc). github.com/autonomous-ai/openharness/tree/main/cl… Oh wait.. you mean to do a PR on Hermes repo?

@trevin @t3dotcodes create a pull request. we'll review and merge! autonomous-ai/openharness

We built an open-source orchestrator to manage coding agents without terminal babysitting The way a lot of us use coding agents is shifting fast from single prompt-and-response chats to having multiple agents running tasks in parallel across different terminals. Leaving Claude Code to ship a feature while Hermes runs tests sounds awesome on paper. In reality, the second you step away to do something else, your mind is still buzzing around your terminal, wondering if an agent finished or if it’s just sitting there hanging on a permissions prompt. That constant mental tether is the exact friction we wanted to get rid of. I work at Autonomous, where we’ve been building Harness, an open-source app that unifies agents running across different machines into one place, accessible via desktop, CLI, or phone. autonomous-ai/openharness Under the hood, every agent runs natively in its own tmux pane and git worktree to keep your main workspace clean. You get pinged only when an agent actually needs human input, so you can let them run async without watching the screen. We’re also playing around with a small physical desk display that shows agent states at a glance without hogging screen real estate. How are you guys currently tracking long-running agent tasks when you step away from your desk? TMUX scripts or something else?

@zalberico love this! have you checked out our open source harness device? autonomous-ai/openharness