Agent Deck: A Terminal Mission Control for Your AI Coding Agent Fleet

What Problem Agent Deck Solves

Running Claude Code on ten projects, OpenCode on five more, and another agent in the background is a real workflow for many developers. The README describes Agent Deck as "mission control for your AI coding agents," and that is accurate. Without a tool like this, you would need separate terminal windows or tmux panes for each session, remember which project runs which agent, and manually track costs. Agent Deck gives you a single terminal UI that lists every session as running, waiting, or done, and lets you switch between them with one keystroke. It targets developers who juggle multiple AI agents across multiple projects, not someone running a single Claude instance.

The Core Mechanism: TUI, Registry, and Conductor

Agent Deck runs as a terminal user interface written in Go. The README shows commands like `agent-deck add . -c claude` to add the current directory with Claude, and `agent-deck` launches the TUI. The TUI maintains a registry of sessions, which you can group, search, archive, and fork. Forking creates a new session that inherits the parent conversation history through the tool's native fork support, so you can explore different approaches without losing context. The conductor is a separate feature: a session that supervises all other sessions. You set it up with `agent-deck conductor setup work --description "Work fleet"` and then start it. The conductor answers routine questions, escalates interesting ones to your phone via Telegram, and monitors workers so a `waiting` session does not rot. The data flow is straightforward: the TUI talks to your configured agent CLIs, stores session state locally, and the conductor reads that state to act on it.

Getting Started: Installation and Quick Start

Installation is a one-liner: `curl -fsSL https://raw.githubusercontent.com/asheshgoplani/agent-deck/main/install.sh | bash`, then run `agent-deck`. Homebrew and `go install` are also available. The quick start shows the essential commands: `agent-deck add . -c claude` to register a session, `agent-deck session fork my-proj` to fork, `agent-deck session send my-proj --message-file task.md` to send a multiline prompt, and `agent-deck mcp attach my-proj exa` to attach an MCP server. The TUI has a full set of key shortcuts: `n` creates a new session, `f` forks quickly, `m` opens the MCP Manager, `s` opens the Skills Manager, `$` opens the Cost Dashboard, and `/` searches. One notable change in v1.9.55: in the new-session dialog, Enter now advances to the next field instead of submitting; you must press Ctrl+S to create the session. You can restore the old behavior with `[ui].new_session_enter_advances = false` in the config. This is a small but important detail for anyone upgrading.

MCP and Skills Management Without Config Editing

A major selling point is the MCP Manager. You define MCP servers once in `$XDG_CONFIG_HOME/agent-deck/config.toml` (default `~/.config/agent-deck/config.toml`), then toggle them per session with the `m` key, cycling scope between LOCAL and GLOBAL with Tab. Agent Deck restarts the session automatically when you attach or detach an MCP server. This removes the need to edit tool-specific config files for each project, which is a real time-saver. The Skills Manager works similarly for Claude skills: press `s` to attach or detach skills per project, with a managed pool workflow. The README shows `agent-deck skill attach my-proj docs --source pool --restart` as a CLI example. This is a concrete mechanism that ties directly into how the tool manages external agent configurations.

The Conductor: Phone-Controlled Supervision

The conductor is the most distinctive feature. The README claims you can go from zero to a Telegram bot that watches every Claude session in five minutes. The setup involves creating a Telegram bot via BotFather, getting your user ID, then running `agent-deck conductor setup work` and `agent-deck session start conductor-work`. From then on, you can message the bot `/status` to get updates. The conductor escalates interesting questions to your phone and prevents `waiting` workers from being ignored. This is a clever approach to remote monitoring, but it depends on external services like Telegram. The README mentions support for Slack and Discord as remote channels, and a `docs/WATCHER-SETUP.md` for adding external event triggers like GitHub events or email. The trade-off is clear: you are adding a dependency on a third-party messaging platform and a separate daemon, which increases operational complexity.

Limitations and When It Is the Wrong Tool

Agent Deck is not for everyone. If you run a single agent on a single project, the overhead of installing, configuring, and learning the TUI is not justified. The tool also has specific version requirements: Codex forking requires a codex CLI with `codex fork <session-id>` support, verified with `codex-cli 0.137.0`. If your CLI does not support that, forking will fail. The README warns about the new-session dialog change, which shows that behavior can shift between releases and break muscle memory. The conductor depends on a separate setup and external messaging services, which may be overkill or a security concern for some teams. The README also notes that the project is actively seeking maintainers, which suggests the bus factor is a consideration. If you value stability over rapid feature development, this is a point to weigh.

Alternative: Native Tool-Specific Session Managers

The obvious alternative is to use each agent's native session management. Claude Code has its own `--resume` and `--continue` flags, and Codex has `codex fork`. These tools are designed for their specific agents and do not require a separate TUI. The difference is that Agent Deck provides a unified interface across multiple agents, while native tools offer deeper integration with a single agent. For example, Claude's native session resume is likely more reliable than a third-party wrapper, because it does not depend on a separate project's maintenance. However, native tools do not give you a single view of all sessions, nor do they offer MCP toggling across agents, cost tracking, or a phone-based conductor. If you only use one agent, stick with the native tool. If you need cross-agent management, Agent Deck is the only tool in this material that provides it.

Maintenance and License Considerations

The project is licensed under MIT, which is permissive and allows commercial use, modification, and redistribution, with the only requirement being to preserve the copyright notice. The repository is actively maintained, with recent releases v1.15.0, v1.14.0, and v1.13.0 pushed in August 2026, and the README states that PRs are validated within about a day. This suggests a fast release cycle, which is good for getting fixes but also means frequent changes to behavior, as seen with the Enter key change. Upgrade cost is real: you need to read release notes to catch such changes. Uninstalling is supported with `agent-deck uninstall` and `agent-deck uninstall --keep-data` to preserve sessions. The README mentions a troubleshooting guide for uninstall, but the details are truncated. The project is not archived, and the maintainer is actively seeking co-maintainers, which is a positive sign for long-term viability, but it also means the bus factor is a concern.

Editorial conclusion

Adopt Agent Deck if you run multiple AI coding agent sessions across projects and want a single TUI to switch, fork, and monitor them, especially if you need MCP toggling or a phone-based conductor. Skip it if you use only one agent on one project, as the setup and learning curve outweigh the benefits. Before adopting, verify that your specific agent CLI versions support the required features, such as codex fork for Codex sessions, and test the conductor with a local channel first to ensure it meets your workflow needs.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes