mmux vs Agent Deck

What is Agent Deck?

Agent Deck describes itself as "mission control for your AI coding agents" and a "terminal session manager for AI coding agents" [1]. It is an open source project hosted at github.com/asheshgoplani/agent-deck, written in Go with Bubble Tea, MIT licensed, and runs on macOS, Linux and Windows through WSL [1][4]. You install it with a shell script, Homebrew or go install, and it needs tmux. The repository mentions no paid tier. Usage telemetry is opt-in and off until you agree [1].

The TUI is a grouped list of sessions with a preview pane that shows the last lines of the selected session and refreshes every two seconds [2]. Each session is a tmux session; Enter attaches to it and Ctrl+Q detaches back to the list. It knows about Claude Code, Gemini CLI, OpenCode, Codex, Copilot, Cursor, pi and several more, with different levels of integration per tool, and any other command can be added as a custom tool [1].

Around that list it adds conversation forking, an MCP manager with a shared socket pool, a skills manager, Docker sandboxes, cost tracking, a web UI (agent-deck web), remote instances over SSH, and "conductors": long-running agent sessions that watch the others and can be reached through Telegram or Slack [1][3][5]. The project releases often; v1.16.28 came out on 2026-10-10 [6].

Where Agent Deck is better

• Forking conversations. f forks the selected session through the tool's own fork support, and by default puts the fork in a new worktree with the parent's uncommitted changes [1]. mmux has no fork.
• Agent configuration tooling. MCP servers and Claude skills can be attached per session or per project from the TUI, a socket pool shares MCP processes across sessions, and per-group Claude accounts are supported [1]. mmux leaves all of that to each agent's own config.
• Remote and phone access. A browser UI, Telegram and Slack bridges through a conductor, and management of agent-deck instances on other machines over SSH [1][3]. mmux is reached over SSH like any terminal program and has no web UI.
• Docker sandboxes. A session can run in a container with the project mounted, and agent-deck try runs a one-shot sandboxed task [5]. mmux runs agents directly on the host.
• Cost tracking. A dashboard ($) shows token spend by day, session and model, with optional budget limits [1]. mmux does not track cost.
• More worktree options. Sparse checkout for large monorepos, Jujutsu workspaces, bare-repository layouts, and setup plus teardown scripts per repository [1].
• Many projects in one list. Groups, fuzzy search, a global search across Claude conversations, archiving and profiles are built for dozens of sessions at once [1][2].

Where mmux is better

• The agent stays on screen. The selected agent, terminal or process runs live in the main pane next to the sidebar, so you type into it without leaving the overview. Agent Deck shows a preview and switches to the full tmux session when you attach.
• Git is part of the window. The git panel stages files and folders, commits with a suggested message, pushes, pulls, switches branches, shows diffs, and can schedule a commit for later. No separate tool needed.
• Dev processes are first class. Processes have start, stop and restart keys, an optional stop: teardown such as docker compose down, and mmux read for their logs. Across worktrees only one checkout runs a given process at a time, so ports stay the same.
• Worktrees as project boxes. w cuts a worktree with a generated name that opens as its own box with its own agents, terminals and git panel. M merges it back, and finished worktrees (merged or pushed, clean, idle) are removed after 30 minutes by default.
• Quit and come back. After q, a crash or an update restart, reopening the directory restores the agents and terminals that were open, and Claude, Codex, Pi and Grok resume their conversations.
• Desktop notifications over SSH. Bells and notification escapes from any session become desktop popups on the machine your terminal runs on, because mmux sends them as escape sequences through tmux.
• Small surface. One binary, a YAML file per project and an optional global one. Fewer features also means less to configure.

When to use which

Moving from Agent Deck to mmux

1. Install mmux. It needs tmux, which you already have for Agent Deck.
2. In each project, run mmux init. The wizard offers presets for Claude, Codex, Pi, Gemini, Amp, opencode and Grok, including their auto-approve modes. Other tools go under agents: in ~/.mmux/config.yaml with a cmd and args.
3. If you used .agent-deck/worktree-setup.sh, move it to worktrees: setup: in mmux.yaml, and list gitignored files to copy under worktrees: copy:.
4. Add dev servers as processes: in mmux.yaml instead of running them in shell sessions.
5. For several repositories in one sidebar, run mmux init workspace in their parent directory.
6. Scripts that called agent-deck session send can use mmux send and mmux wait.

curl -fsSL https://mmux.org/install.sh | sh
cd ~/some/project
mmux init
mmux

Your conversations live in each agent's own history, not in Agent Deck, so nothing needs exporting. The two tools can also stay installed side by side: Agent Deck names its tmux sessions agentdeck_* [1], and mmux sets its tmux options per session, never globally.