LAYER 1The client
The desktop app, a tablet browser or the Android app. The client stores host records, credentials and the decision log. It does not run agents.
AgentMux manages AI coding agents running on remote hosts and on this computer, including Claude Code, Codex, Gemini CLI, Grok CLI and Aider. Each agent runs in a tmux session on its host, so closing the client, suspending the laptop or losing the network does not interrupt the task. It also works from a tablet or a phone.
AgentMux does not host the agent process. It connects to a tmux session on the host, where the agent runs, so the task does not depend on the client being online.
The desktop app, a tablet browser or the Android app. The client stores host records, credentials and the decision log. It does not run agents.
One SSH connection per host. Terminals, commands and file transfers share it, so opening several terminals requires one authentication. Local hosts connect directly without sshd.
The agent runs in a tmux session on the host. The session keeps running when the client closes, the laptop sleeps or the network drops. Reopening the client reattaches to the same pane with its scrollback.
AgentMux does not host agent processes. It connects to tmux sessions on the host, so whether a session is alive does not depend on the client.
Agents keep running when the client disconnects, the laptop sleeps or the network drops. Reopening the client reattaches to the same pane with the full scrollback. Long tasks no longer depend on an unbroken SSH connection.
Each agent pane is a full terminal with colour, mouse, selection and search. You can type into it, correct the agent and hand control back without restarting the task.
Split the terminal area into up to nine panes — a 3×3 wall across hosts, each independently interactive. Drag the seams, double-click a tab to zoom, and the layout reflows with the window and is restored on next start.
Send one instruction to any selected agents. Each agent returns a delivery receipt, so the result of a batch operation can be checked per agent.
A local Ollama model can work on a goal, for example finding out why an agent is stuck, one tool call at a time. Any write operation pauses for your approval; scheduled checks are read-only.
Project facts, preferences, and agent activity are indexed for retrieval by wording or by meaning. Embeddings are computed locally; nothing leaves the machine. Credentials matching known patterns are redacted before storage.
The same binary runs headless, and a tablet browser then shows the complete AgentMux: terminals, agents, toolbox and file browser. Android also has a standalone app with the core built in; SSH connections start from the device, so no always-on machine is needed.
Agents in the tree used to show one state: running. AgentMux now reads each pane periodically and classifies it: asking a question, finished and waiting for review, or idle at the prompt. Agents that need attention are highlighted in the tree until someone looks at them.
AgentMux checks for a new release at startup and every six hours, downloads it, verifies the sha256, replaces itself and restarts, rolling back on failure. The full configuration can be exported as an encrypted file and imported on another computer without overwriting existing entries.
All images below are application screenshots.
Attach to any agent's tmux pane: full colour, mouse reporting, search and selection. The side panel shows process state, recent output, and lifecycle controls.

Send the same instruction to any selected agents across projects and hosts. Each agent returns a delivery receipt, so every delivery can be confirmed.

The install panel inspects a host first and offers only the agent CLIs and runtimes it can actually support, stating why others are unavailable. Installation runs inside tmux, so a dropped connection can't leave a partial package tree.

The default palette follows the Apple system colours, with six more themes including Nord, Solarized and a light mode. Information density is tuned for long monitoring sessions.

Besides the desktop app there are two more ways to use AgentMux: headless serve mode makes any browser a full client, and Android has a standalone app with the core built in. Both run the same front end on the same core, with the same features as the desktop app.
Releases ship a dedicated server build, agentmux-server-linux-{amd64,arm64} — fully static, with no GTK, no webview and no display needed. Copy it to any Linux server and run it. The desktop build enters the same mode with --serve. The first start generates an access token and prints it in the log (kept as serve-token in the data directory; AGENTMUX_TOKEN overrides it).
Point the tablet at http://host:8642, enter the token once, and everything works — terminals, agents, the toolkit, file browsing. “Add to Home Screen” (Safari on iPad, Chrome on Android) installs it as a standalone app. Closing the browser stops nothing, exactly like closing the desktop window: the agents live in remote tmux.
Every release attaches agentmux-android.apk (Android 8.0+ · arm64) with the same core the server build ships, embedded and started by a foreground service — foreground because Android freezes background processes at screen lock, and the core holds the SSH connections. The layout was reworked for small screens too: under 768px the side panels become overlay drawers, and a bottom navigation bar puts the tree, the command palette, the detail panel and settings where a thumb actually reaches, safe-area inset respected.
The desktop app, headless serve and the Android app share one core. The difference is where the core runs.
| Desktop app | Headless serve | Android app | |
|---|---|---|---|
| Core runs on | Your computer | Any Linux box | The device itself |
| Needs an always-on machine | No | Yes (the server) | No |
| How you use it | Desktop window | Any browser: iPad, Android tablet, phone | Native app, kept alive by a foreground service |
| Where state lives | Local SQLite | On the server, shared by every client | On the device |
| Authentication | System keychain | Access token; put HTTPS in front for the public internet | On the device |
| Download | agentmux-macos / windows / linux | agentmux-server-linux-{amd64,arm64} | agentmux-android.apk |
Besides tablet and phone support, the August release includes the following.
A terminal that lost its transport is not over — the tmux session on the far side is still running. The pipe is rebuilt with the same shell id, the same scrollback and the same pane, backing off from one second to fifteen and leaving the manual button after about five minutes. Two keepalives, TCP and SSH, answer different questions; one-shot command terminals are exempt, because re-running an install is a side effect nobody asked for.
Some work is a screen. Right-click a host and 3389 or 5900 is dialled through the SSH connection that host already has, forwarded to a port here and handed to whichever viewer this computer has — Remote Desktop Connection, Screen Sharing, Remmina. A desktop listening on its own loopback needs nothing opened to the network.
The app checks the release feed shortly after launch and every six hours after. A newer version raises a one-line banner under the title bar: upgrade and restart, release notes, or later. Upgrading downloads with a progress bar, verifies the published sha256, swaps the running executable, and rolls back to the build that was running if the swap fails.
Hosts, folders, projects, workspaces and agent definitions go into a single passphrase-encrypted file (Argon2id + AES-256-GCM) and open on the other side. An import never overwrites: a host at the same address, a project with the same name, a path on the same machine — each is left exactly as it is, so importing twice adds nothing.
The metrics panel now reports CPU model, DIMM specs, physical drives and graphics adapters alongside utilisation. Those are static, so they ride a one-shot read cached for the session rather than the three-second ticker.
One-click install now covers Grok CLI, Docker (six methods, retrying against a mirror where the network needs one, then starting the daemon and adding the user to the group) and Ollama. The interface adds Traditional Chinese — a full catalogue in Taiwan terminology, not a character-for-character conversion.
The constraints below are design decisions, not settings. Most of them cannot be turned off.
Passwords and key passphrases are encrypted with AES-256-GCM before storage; the key lives in the OS keychain (Keychain, Credential Manager, Secret Service). If the keychain is unavailable, a 0600-file fallback is reported in the status bar.
Host keys are pinned on first connection. Any subsequent mismatch aborts the connection with an explanation — a man-in-the-middle swap cannot pass silently.
No secret crosses into the UI process. The interface learns only whether a secret is set — never its value.
The orchestrator can only call a fixed whitelist of tools, each carrying a risk tier fixed at declaration. Execution passes through a single gate combining tier, host trust level, and the run's trigger.
Scheduled patrols are refused every non-read tool unconditionally — and that restriction is not configurable. A patrol can report a stalled agent, but has no authority to act on it.
Every step of every run is recorded — including proposals that were refused, rejected, or left unanswered. Remote output enters the model marked as data; instruction-shaped text raises a flag on the approval card.
The architecture of AgentMux follows four principles.
AgentMux connects to tmux sessions on the host and does not host agent processes. Quitting the client does not affect running tasks.
One binary and one SQLite database. There is no account, daemon or backend service, and host information stays on your computer.
The orchestrator investigates and proposes; any change to a host needs your confirmation. Unattended checks are limited to read-only operations.
Embeddings are computed locally and secrets are not passed to the UI process. Agents connect to model providers directly from the host; AgentMux does not proxy that traffic or read API keys.
Migrations, refactors and large test runs that take longer than a working day or a single network connection.
Agents running on three to fifty hosts, shown in one tree with each agent's state and latest output.
One operation sends an instruction to many agents and collects a delivery receipt from each.
Set up a new host from the install panel instead of assembling install commands by hand.
Scheduled read-only checks that report stalled agents without making changes.
Planning, embeddings and memory run locally; agent model traffic is not proxied and API keys are not read.
Unix-like remote hosts need only tmux and an SSH account — AgentMux can install tmux where it's missing; a remote Windows host just needs OpenSSH Server. This computer is even simpler: no credentials at all.
Read the full docs→Remote: address, user, and one of ssh-agent / key / password, with jump hosts supported. Local: nothing but a name.
A working directory on that host. Or browse the host's files and add the directory in place — its name, path, and host are already known.
A name and the command that starts it. Hit Start and it's running inside tmux on the host.
claude --dangerously-skip-permissionsAttaching to an agent, opening a shell, installing a CLI, changing theme — the fastest route to all of it.
Ctrl / ⌘ + KAll produced by GitHub Actions from a tagged commit, each file with a .sha256 alongside.
Any Linux · amd64 and arm64 · no GTK, no webview, no display
agentmux-server-linux-amd64.tar.gz
agentmux-server-linux-arm64.tar.gz
Android 8.0+ · arm64 · embeds the full core, runs standalone
agentmux-android.apk
Download ↓xattr -dr com.apple.quarantine);
on Windows, "More info → Run anyway". Downloading with
curl -L -O avoids the macOS quarantine attribute.
No. Objectives, prompts, and repository conventions remain yours. The orchestrator investigates and proposes; any action that modifies a host executes only after your explicit approval.
Remote Unix-like hosts (Linux / macOS) need a POSIX shell, tmux, and SSH (AgentMux can install tmux where it's missing); a remote Windows host just needs OpenSSH Server enabled — sessions are hosted by AgentMux's own session daemon, deployed over SFTP on first use. This computer qualifies directly on Linux and macOS; on Windows the same machine offers two local hosts — the WSL distribution (where tmux lives) and native Windows (PowerShell, persisted by the same daemon, so closing the window does not stop native work either).
No. Only local orchestration and semantic memory search need Ollama (one chat model plus one embedding model). Without it, everything else is fully functional.
No — deliberately. Agent traffic goes straight from the host to whichever provider the agent CLI is configured for. AgentMux does not proxy it and does not read the keys, so it can neither report cost nor enforce a budget.
One operator per installation. State, credentials, and the decision log live on the workstation — no shared server, no team account, no central audit sink. Two people working the same fleet see the same tmux sessions but keep separate histories.
Never. Removing a workspace or agent record only removes the local record — the tmux session keeps running. The one control that destroys a running session is named Kill, and it confirms first, stating what is lost.
Two ways. Run the headless server on any Linux machine — the dedicated agentmux-server-linux-{amd64,arm64} build, or the desktop build with --serve — then open http://host:8642 on the tablet, enter the token once and “Add to Home Screen” for a standalone app. That is the iPad route. On Android you can instead install agentmux-android.apk, which embeds the full core and needs no server at all.
In serve mode the core runs on one of your machines and the tablet is a client, so phone, tablet and desktop see the same state. The APK puts the core on the device itself: SSH runs from the phone, configuration and keys stay local, and no always-on machine is involved. The foreground service is deliberate — without it Android freezes the process at screen lock and cuts every SSH connection.
It authenticates with an access token alone, generated on first start and kept as serve-token in the data directory. That is enough on a LAN or over a VPN; for public access, put an HTTPS reverse proxy in front rather than exposing plain HTTP.
A single binary with local SQLite storage and no server. MIT licensed, ready to download.