Headless servers, GPU rigs, a NAS, a Raspberry Pi, the spare Mac Mini in the cupboard. Give Claude, ChatGPT, Codex, or any MCP/HTTP client a shell on your hardware — and start work that keeps running long after the chat ends.
Free covers AI access to up to 10 saved machines. Outbound-only: no exposed SSH, open ports, or VPN. Commands and their output pass through the relay in memory only — never logged, never stored — and your files stay on the machine unless you transfer one.
Install the app, give your AI the machine code, and ask for what you need. The app connects outbound, so there are no firewall changes.
On Linux — a cloud server, home server, or Raspberry Pi — paste one line. On Mac or Windows, install the desktop app. You get a stable code like AIC-7K3P-WX9M-RTBN.
Give the code to ChatGPT, Claude, or another AI assistant, then ask in plain language. It runs the work on that machine and returns the result.
Sign in for aliases and a command center. Free includes up to 10 usable saved devices; Pro is $49/month and enables all saved devices (up to 100**) plus file transfer. You do not need an account to try ordinary commands during a code's first hour. Compare plans →
** 100 machines is a technical ceiling, not a policy limit. Need more? Get in touch — we'll sort it out.
Start it as a detached job on your own machine. The MCP or REST call returns a jobId immediately; the build, download, render, data pipeline, or training run continues without keeping the client, laptop, or network connection alive. Close the chat and check it later.
Ask in plain language or call remote_job_start. Come back from another chat, client, or phone to inspect state and retained output. The same lifecycle is available over POST /api/v1/jobs and related REST endpoints.
gpu_index to prevent collisions.Yes — and it needs no special tool. Because remote_job_start runs any command, the agent you are already talking to in ChatGPT or Claude can launch the coding agent installed on a different box, then walk away and collect the result.
A cloud model with no filesystem starts claude -p "…" or a codex run as a detached job on the machine where the checkout, the credentials, the toolchain, and the GPU actually live. The call returns a jobId at once, so the conversation is never blocked by a twenty-minute refactor.
remote_job_status tells you whether it is still running or what it exited with; remote_job_logs pages through the coding agent's own transcript in bounded slices.remote_pull brings the diff, the build artifact, or the report back as a real file instead of squeezing it through command output.No magic tool involved. There is no ask_agent endpoint: the mechanism is remote_job_start plus whichever agent CLI you have already installed on that machine.
The first connection to a box is always slow: the model pokes around to find out what is installed, where the projects live, and how anything is built. AI Commander gives it somewhere to write that down — so the second session starts informed instead of re-exploring.
When you are signed in and the machine is online, session_status hands back the path of a note file kept on that machine for your account — people sharing a box keep separate notes. The AI reads it before it starts poking around, and appends what it learned: installed toolchains, where projects and data live, how to build and run them, and this box's quirks.
remote_exec, and the contents never leave your machine except as the answer to your own command.With Pro, download a checkpoint, report, log bundle, or rendered image from the machine, or upload a config, small dataset, or release artifact. Free and anonymous callers cannot start file transfers.
On Pro, remote_pull copies one regular file from the remote machine into a temporary relay blob and returns a download link. Each file can be up to 100 MiB; package a directory into one archive first.
Upload your bytes, then remote_push writes that blob to a destination path on the machine. Uploading and pushing require a signed-in Pro account.
Pick your platform. The app gives you the machine code you will use to connect.
Menu-bar app, everything included · Apple Silicon (M1–M4) · Intel (x64)
Tray app, everything included · Windows x64
Signed release installer verified before sudo · x86_64 & arm64 · OpenSSL 3 required
Buttons above use mutable /dist/latest/* paths for convenience.
For stronger assurance, prefer the immutable release directory and checksum manifest on GitHub.
Desktop PKG/DMG/ZIP installers are signed with Apple Developer ID and notarized.
They do not ship detached Ed25519 signatures like the Linux agent.
Download from the GitHub Release, then verify with
shasum -a 256 -c SHA256SUMS after extracting the manifest alongside the artifact.
Gatekeeper validates code signing when you open the installer.
The NSIS .exe is Authenticode-signed (Azure Trusted Signing).
There is no detached Ed25519 signature. Download from the GitHub Release and verify with
certutil -hashfile AICommander-Setup.exe SHA256 against the matching line in
SHA256SUMS. Windows SmartScreen checks the embedded signature at install time.
Use the signed installer flow in install docs:
verify install.sig with the pinned public key before sudo.
Direct binary downloads from /dist/latest/agent-linux-* include matching
.sha256 and .sig files, but the installer pins an immutable
/dist/v/<ver>/ set.
Connect Claude, ChatGPT, Codex, or another AI client. Signing in is free and keeps your machines available by name. For a one-hour account-free trial, add ?anonymous=1 to the URL. Most AI tools need just one command:
claude mcp add --transport http aicommander https://aicommander.dev/mcp
codex mcp add aicommander --url https://aicommander.dev/mcp
opencode mcp add aicommander --url https://aicommander.dev/mcp
pi install npm:@aicommander/mcp
Or give a capable AI agent the site and machine code in plain language:
use aicommander.dev to connect to AIC-XXX
Using Cursor, Windsurf, ChatGPT, or Claude Desktop instead? All connection guides →
Each option lets your AI run commands and jobs or transfer files. Pick the one that fits the tool you already use.
The simplest path for Claude, Codex, opencode, Cursor, ChatGPT, and other MCP clients. Run one command and you're connected. Setup →
Prefer plain web requests? Send the code and your command to a simple web address. Anything that can call a URL — a script, a scheduled job, or a chatbot — can drive a machine. API docs →
Drop a ready-made Skill into any AI agent that supports skills. Then mention a machine code and it works. Install the Skill →
A direct way for your AI to work on a real machine: install a small app, provide the machine code, and say what you need. Great for checking logs, running tests, restarting an app, or managing a machine with no screen — with nothing left open or exposed.
The category has filled up fast. Here is where we win, and — because you will find out anyway — where we do not.
| AI Commander | Desktop Commander | ManageLM | Chaterm | SSH-MCP plugins | Rented cloud GPU (e2b, Modal, RunPod) |
|
|---|---|---|---|---|---|---|
| Runs on hardware you already own | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ rented |
| Headless box with no GUI or desktop session | ✓ command-first | — | ✓ servers | ✓ over SSH | ✓ over SSH | ✓ |
| Detached jobs that survive the chat ending | ✓ built in | — | cron schedules | — | tmux/nohup by hand | ✓ in their cloud |
| NVIDIA GPU discovery + exclusive per-job reservation | ✓ gpu_index |
— | — | — | ✗ | ✓ on rented GPUs |
| Fleet of machines addressed by alias | ✓ aliases | — | ✓ server groups | ✓ SSH host list | one per config entry | — |
| Nothing exposed inbound — outbound-only agent | ✓ | — | ✓ outbound WebSocket | ✗ needs reachable SSH | ✗ needs reachable SSH | — |
| Your files stay on the machine that runs the work | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ their infrastructure |
| Per-machine notes the AI reads and writes | ✓ per account | — | — | local memory | — | — |
| Agent-to-agent: start another machine's coding agent as a job | ✓ remote_job_start |
— | — | — | no job lifecycle | inside their sandbox |
| GUI control — clicking and typing on a desktop | ✗ screenshots only* | — | — | ✗ terminal | ✗ terminal | ✗ |
| Self-host the control plane yourself | ✗ hosted relay only | — | ✓ Docker Compose | ✓ GPLv3 | ✓ open source | ✗ |
| Teams, roles, and audit (RBAC) | ✗ not today | — | ✓ teams + RBAC | — | ✗ | — |
| Entry price | Free — 10 machines Pro $49/mo — up to 100** |
MCP server free (MIT) hosted app sells AI usage credits from $20/mo |
Free — 10 agents | Free (GPLv3) · Pro $99/yr | Free (open source) | Metered compute |
Competitor claims verified against each vendor's own site on 26 August 2026: desktopcommander.app, managelm.com, chaterm.ai. A dash means we could not verify the capability either way from public material and would rather leave it blank than guess. “SSH-MCP plugins” describes the common shape of the many community MCP servers that wrap an SSH client, not one product. Products move; tell us if something here is out of date.
* Screenshots need a desktop session — they are available on the macOS and Windows desktop apps only, after the machine owner enables the time-limited screen-sharing grant. macOS also requires its separate Screen Recording permission.
** 100 machines is a technical ceiling, not a policy limit. Need more? Get in touch — we'll sort it out.
The classics still work. They just were not built for an agent to drive.
| AI Commander | SSH | TeamViewer | Tailscale + SSH | |
|---|---|---|---|---|
| Built for AI | ✓ built in | manual | screen-first | ✗ |
| Works with Claude, ChatGPT, Codex | ✓ built in | ✗ | ✗ | ✗ |
| No open ports needed | ✓ | ✗ needs a port | ✓ | ✓ |
| No screen needed | ✓ headless | ✓ | ✗ screen needed | ✓ |
| Work that survives the chat | ✓ built in | tmux/nohup by hand | ✗ | tmux/nohup by hand |
| Works behind a firewall | ✓ | extra setup | ✓ | ✓ |
| What you get back | Text, files & screenshots* | Text you can read | A screen to watch | extra setup |
| Built-in file transfer for AI | ✓ pull & push on Pro | SCP/SFTP separately | manual file transfer | SCP/SFTP separately |
Long work doesn't have to be babysat. A training run, a big build, or an overnight dataset job is started as a
detached job: it keeps running on your machine after the chat ends, the laptop sleeps, or the network drops.
You can ask how it's going later — from another chat, another client, or your phone — read the log, or cancel it.
It carries on when the agent itself restarts too: on Linux each job is launched into its own transient
systemd scope, outside the agent service's control group, so restarting or upgrading the agent
leaves it running — provided the box runs systemd and the agent runs as root, and without both an
upgrade still stops running jobs. On Windows a job survives the agent process dying by itself, but an installer
that stops the app and everything it started still ends it — assume an unattended auto-update is one unless that
machine's installer is known to do otherwise.
With ssh you'd wrap the same command in tmux or nohup, keep the session alive, and reconnect to find out what happened.
Start ML training on your own GPU →
AI Commander gives your AI real access to your machine, so security comes first. It is safe by default and explicit about the limits.
The app only reaches out — nothing on your machine waits for incoming connections. There's no open door to find or attack, and it works behind firewalls untouched.
We never keep your codes or keys in readable form. Even if our database leaked, there would be nothing usable in it — the only readable fragment is a two-character label for your linked machines, which cannot be used on its own. You can revoke a key at any time.
Access refreshes on its own and quietly expires. Keys go dormant until you sign in again, so a forgotten one can't keep working forever.
Your commands and their text results are never logged or saved. Files you explicitly transfer are the exception: access ends after 24 hours, then an hourly, retrying sweep removes the inaccessible bytes.
A brand-new code works for anyone for one hour, so it's easy to get started. After that, only approved accounts can use it — until you choose to reset and clear access.
Whatever a machine sends back is treated as plain results to show you, not as new orders to follow — so a sneaky line in a log is far less likely to fool the AI.
Need unattended access for a script or scheduled job? A service token runs as a non-root user, on one machine, with only the exact commands you allow — and no shell, so nothing can be injected. Tokens can be long-lived and revoked at any time.
** 100 machines is a technical ceiling, not a policy limit. Need more? Get in touch — we'll sort it out.
Each has a free tier.