The useful layer above tmux
Agent Deck solves a problem that appears only after AI coding becomes parallel. One agent is easy to watch in a terminal. Ten agents spread across repositories, worktrees, accounts, and remote machines are not. Agent Deck puts those sessions in one terminal interface, marks whether each is running, waiting, idle, or in error, and attaches to one with a keystroke. Underneath, tmux keeps the processes alive.
That basic dashboard is the project's best feature. Groups organize sessions, fuzzy search finds them, archived sessions retain conversations, and a status-bar list highlights workers that need input. The registry outlives the current terminal, so a developer can stop treating every agent window as disposable shell state. Agent-specific resume and fork support makes the experience better than a handmade tmux script when a supported CLI exposes the required conversation controls.
The strongest integration is Claude Code. Gemini CLI and Codex also get MCP handling and status features, while OpenCode, Copilot, Cursor, Hermes, DeepSeek Harness, and custom commands have varying subsets. This is an honest support matrix, not a promise that every terminal tool behaves identically. For Codex, turn-level status depends on installing a notify hook for each Codex home. Without it, waiting and completion state cannot converge reliably.
Parallel work has practical guardrails
Git worktree support is the feature most likely to prevent real damage. Agent Deck can start a session on a new branch in an isolated worktree, copy selected ignored files through .worktreeinclude, and run bounded setup and destruction scripts. Sparse-checkout inheritance helps large monorepos avoid materializing an entire tree for every worker. Forking can carry conversation context, working-tree changes, branch state, and Docker preference into a new line of attack.
These conveniences still deserve review. A copied .env remains a secret in another directory. Setup scripts run against newly created worktrees, and destruction scripts do not stop removal when they fail. Finishing a worktree can merge and clean up state, so users should understand the exact command before turning it into an agent action. The project documents these semantics well, which makes cautious adoption possible.
Docker mode limits access to the wider host, but it does not make the repository read-only. Agent Deck bind-mounts the project read-write because the agent must edit it, and shares host tool authentication into the container to avoid another login. On macOS it can extract Keychain credentials. That is useful isolation from unrelated host files, not a safe place to run instructions that should not touch code or use your agent account.
The control plane keeps expanding
MCP and Claude skill managers let a user attach configured tools to a project without editing each agent's configuration by hand. A socket pool can share MCP processes across sessions. Cost views read transcripts or tool output, group spending, and expose budget settings. Remote support discovers Agent Deck installations over SSH and shows remote sessions beside local ones.
Then there is Conductor, a persistent Claude or Codex session that supervises other sessions, answers routine questions, and escalates through Telegram or Slack. Watchers can wake it from GitHub webhooks, Gmail, generic webhooks, notifications, or scheduled events. This is powerful automation, but it crosses from session organization into a system that can receive outside events and direct coding agents. Each bot, webhook secret, policy file, and credential boundary becomes operational configuration. Start without it unless unattended fleet supervision is the problem you already have.
The web interface has a sensible default: it listens on loopback. A non-loopback bind is refused without a bearer token because the API and terminal form a remote-code-execution surface. For headless use, the README recommends a permission-checked token file so the secret does not appear in process arguments. Read-only mode is available for observers. These are good defaults, but putting the service behind a public reverse proxy still demands normal transport security, access control, and patching.
Active development cuts both ways
Agent Deck v1.13.0 was released on August 16, 2026, and the repository was pushed again that day. Issues and pull requests were still moving on August 17. The open count of 37 combines both types of work, while the recent history shows fixes and features arriving at a remarkable pace. The maintainer actively asks for contributors and one or two co-maintainers, which is both welcoming and a reminder that the surface is large for a small leadership group.
Recent bugs show why version numbers alone should not settle the maturity question. Issue #1956 documented a path where a failed read or parse during an MCP configuration update could erase unrelated Claude settings. It was fixed and closed the same day. Open issue #1960 says the web MCP pane still cannot attach servers to Codex or Gemini sessions because it submits the wrong scope. Another recently fixed issue found restart commands leaving live tmux processes orphaned from Agent Deck's registry. Fast response is reassuring, but users should update deliberately and keep configuration backups.
Some advertised controls also carry explicit caveats. The README calls Gemini, Codex, and MiniMax cost collection through output parsing untested, and says the budget limits are untested. Do not use those totals as accounting records or hard spending controls until verified against the chosen tool and model.
Who should install it
If your current workflow includes several terminal agents, manual worktrees, and periodic hunting for the pane awaiting approval, Agent Deck can replace a surprising amount of glue. Install it, add two local sessions, enable a separate tmux socket, and test stop, restart, resume, and fork behavior before trusting it with a large fleet.
If you want only durable panes, tmux or Zellij is simpler. Claude Squad is the closer alternative for a focused parallel-agent interface. Agent Deck earns its extra complexity when status, session history, worktree lifecycle, and cross-machine visibility all matter. It is best treated as a local operator console first, and an automation control plane only after its permissions and failure modes are understood.