Your AI, working on your own machines.

A small agent on each computer dials out to the relay. Claude, Codex, ChatGPT or any MCP client can then run commands, start long jobs, look at the screen and move files. Five short scenes show how.

  1. 01How it works

    Run the signed installer on a Linux box you own (on a Mac or Windows PC, install the desktop app). The agent dials out to the relay, and aicommander-agent status --reveal shows its session code, such as AIC-7K3P-WX9M-RTBN. Give the code to your AI, and it runs nvidia-smi on the GPU box with remote_exec. Only the output travels back to the chat.

  2. 02Quick command

    ChatGPT asks whether the NAS is running out of space. One remote_exec call runs df and du and returns in under two seconds: 87% full, backups take the most room.

  3. 03Overnight job

    Claude Code starts a fine-tuning run with remote_job_start and gets a job ID right away. The chat closes and the job keeps running. Next morning a new chat checks it with remote_job_status, reads the last log lines with remote_job_logs, and gets a short-lived download link for the adapter with remote_pull.

  4. 04See the screen

    Codex takes a remote_screenshot of a Windows PC whose installer looks stuck, finds a "Restart required" dialog and schedules the restart for 19:00 with schtasks in cmd.exe.

  5. 05Whole fleet

    One question, four parallel remote_exec calls: softwareupdate on the Mac, apt on Linux, winget on Windows. The AI reports which machines need updates.

How the pieces fit

AI Commander has three parts: an agent on each machine you own, the relay at aicommander.dev, and an MCP endpoint that your AI client connects to. The model never gets a network path into your machine. It calls tools, the relay passes each call to the right agent, and the agent sends back the result.

The agent dials out

On Linux the signed installer sets the agent up as a systemd service; on macOS and Windows it comes inside the desktop tray app. Either way it opens a TLS WebSocket to the relay. The connection starts on your side, so the machine needs no open port, no VPN and no SSH key, and it works behind NAT or a home router. See the security model for what the relay can and cannot see.

A stable session code

Each machine gets a code like AIC-7K3P-WX9M-RTBN. The code stays the same across restarts. Treat it like a password: anyone who has it can reach the machine. Without an account, a code works only in its first hour; a signed-in account can use a code of any age until you block that account or reset the code. As the owner you can reset the code, or block an account that should no longer connect. Signed in, you save machines and call them by name. Read more about session codes.

One MCP endpoint for every client

Claude, Claude Code, Codex, ChatGPT, Cursor, Gemini, Grok and other MCP clients connect to https://aicommander.dev/mcp. Commands, jobs, screenshots and file transfers show up as ordinary tools, and every call appears in the transcript. Clients without MCP can use the REST API. The connect guide has the setup for each client.

Jobs that outlive the chat

remote_exec is for commands that finish in seconds or minutes. Long work such as a training run or a big build goes through remote_job_start, which returns a job ID at once. The job keeps running on the machine after the chat closes or the laptop sleeps. Any later conversation can check it, read its logs or cancel it. See detached jobs.

Files and screenshots

remote_pull and remote_push move real files such as checkpoints, images and datasets outside the command output. File transfer is part of Pro; see file transfer for limits. remote_screenshot captures the desktop of a macOS or Windows machine only after its owner turns on Share Screen in the desktop app. It is off by default.

The five scenes

  1. 01 · How it works

    Install the agent, get a session code, and let Claude read the GPU model from a Linux box.

  2. 02 · Quick command

    ChatGPT checks free space on a NAS with a single remote_exec call.

  3. 03 · Overnight job

    A fine-tuning run starts as a detached job, survives a closed chat, and the next morning its logs and a download link for the adapter come back.

  4. 04 · See the screen

    Codex reads a stuck installer from a screenshot and schedules a restart after office hours.

  5. 05 · Whole fleet

    One question fans out to a Mac, a Linux GPU box, a Windows PC and a NAS at the same time.

More real workflows are on the use cases page.

Try it on your own machine

Install the agent on a computer you control, connect your AI client, and ask for a real task. Free for up to 10 machines.

Install the agent Connect your AI Read the docs