AgentsDock Puts Claude Code, Codex and Cursor in One Window — and on Your Phone
AgentsDock calls itself an IDE for agentic AI research, which undersells it a bit. What it actually solves is the problem every person running more than two agents has hit: your agents live on remote boxes, your terminal history dies when you close the laptop, and there is no way to check on a training run from a phone without SSH gymnastics. https://agentsdock.net/
The design decisions are the interesting part. Claude Code, Codex and Cursor run side by side in one workspace, so the harness stops being the thing that dictates where you work. Multiple remote servers connect at once and you switch between them without re-establishing context. Terminal sessions persist through tmux, which means closing a client does not kill anything. Agent outputs — plots, images, videos — render inline in the chat rather than landing as file paths you have to go fetch. And it ships as desktop on macOS, Linux and Windows plus native mobile on iOS and Android, with a self-hosted backend called AgentsServer, so the fleet stays on your machines.
The mobile piece is the part that will either seem trivial or obvious depending on how you work. If your agents run for hours, the useful interaction is not typing code, it is glancing at a loss curve at dinner and queueing the next task. Being able to see a rendered plot on a phone from a training run on a remote box is a small thing that changes the shape of the day.
Be clear about the maturity: 27 stars, 82 commits, desktop at v1.0.0-beta.1 with the stable line at v0.2.12, built by Zhengyi Luo. It claims use by researchers at Carnegie Mellon, Berkeley and NVIDIA. And there is no project-wide license in the repo — third-party components keep their own terms, but the project itself has not declared one, which in practice means you do not have permission to do much with it beyond running it. For a self-hosted tool you are meant to point at your own infrastructure, that is the thing to fix first.
It belongs on the same shelf as [worktrunk, which became a platform because everyone now runs five agents](https://clauday.com/article/b171b084-f6f7-4665-aaef-db92ef695961), and [Herdr going multi-machine](https://clauday.com/article/f99c62c2-f0ff-486e-a796-8cd16b4d2e07). Three different teams, same observation: the unit of work stopped being a session and became a fleet, and the tooling built for one terminal and one repo does not survive the transition.
← Back to all articles
The design decisions are the interesting part. Claude Code, Codex and Cursor run side by side in one workspace, so the harness stops being the thing that dictates where you work. Multiple remote servers connect at once and you switch between them without re-establishing context. Terminal sessions persist through tmux, which means closing a client does not kill anything. Agent outputs — plots, images, videos — render inline in the chat rather than landing as file paths you have to go fetch. And it ships as desktop on macOS, Linux and Windows plus native mobile on iOS and Android, with a self-hosted backend called AgentsServer, so the fleet stays on your machines.
The mobile piece is the part that will either seem trivial or obvious depending on how you work. If your agents run for hours, the useful interaction is not typing code, it is glancing at a loss curve at dinner and queueing the next task. Being able to see a rendered plot on a phone from a training run on a remote box is a small thing that changes the shape of the day.
Be clear about the maturity: 27 stars, 82 commits, desktop at v1.0.0-beta.1 with the stable line at v0.2.12, built by Zhengyi Luo. It claims use by researchers at Carnegie Mellon, Berkeley and NVIDIA. And there is no project-wide license in the repo — third-party components keep their own terms, but the project itself has not declared one, which in practice means you do not have permission to do much with it beyond running it. For a self-hosted tool you are meant to point at your own infrastructure, that is the thing to fix first.
It belongs on the same shelf as [worktrunk, which became a platform because everyone now runs five agents](https://clauday.com/article/b171b084-f6f7-4665-aaef-db92ef695961), and [Herdr going multi-machine](https://clauday.com/article/f99c62c2-f0ff-486e-a796-8cd16b4d2e07). Three different teams, same observation: the unit of work stopped being a session and became a fleet, and the tooling built for one terminal and one repo does not survive the transition.
Comments