Tutorials
Agent Deck Setup: The Commands and Shortcuts That Matter
Beyond the discovery piece: install paths, the shortcut map, MCP and skill attachment, cost tracking and the phone-side conductor — straight from the repo's docs.

Last updated: September 17, 2026 · 7-minute read
Most agent-orchestration tools fail at the boring part: the second week, when the novelty is gone and you just need to know which of your nine sessions is waiting on you. Agent Deck — the Go-based TUI this series covered earlier — is now at 845 stars, still MIT-licensed, still shipping weekly, and its README documents a workflow deep enough to deserve its own guide. This is that guide: every command below comes from the repository's own documentation, verified September 6, 2026.
Install: four paths, one recommendation
Supported platforms are macOS, Linux, and Windows via WSL — note that last one, because native Windows is not on the list. The curl installer is the fastest route:
Homebrew covers macOS handsomely, and Go developers can skip the binary entirely:
Piping install scripts straight to bash deserves a standing caveat: read the script first, or use the brew/go paths where the supply chain is shorter. After any install method, agent-deck alone launches the TUI.
The five commands that structure everything
The README's quick start compresses the whole product into a handful of invocations, and they map cleanly onto how the tool wants to be used:
add registers a directory-plus-agent pair as a session in the deck. fork branches a session's history — the git-worktree-aware version of "let me try that differently". session send with a message file is the batch-work lever: write the task spec once, deliver it verbatim. The mcp attach and skill attach subcommands bind MCP servers and skills to specific sessions, which turns the deck into configuration management, not just a window switcher.
The shortcut map worth taping to your monitor
Inside the TUI, the core loop is four keys, with a second tier worth learning on day two:
The cost dashboard on $ aggregates spend across sessions — the honest number that makes multi-agent fleets feel real. Groups on M matter once sessions cross ten: one group per project keeps the deck readable.
One behavioral change to know before it confuses you: since v1.9.55, pressing Enter in the new-session dialog advances to the next field rather than creating the session — Ctrl+S creates from any field. The dialog also remembers your last-used tool. The README notes the old behavior is one config line away: [ui].newsessionenteradvances = false.
The conductor and the web view
Two surfaces extend the TUI beyond the machine it runs on. agent-deck web serves a local web UI on 127.0.0.1:8420 — sessions visible and manageable from a browser. The conductor goes further: the README describes phone-controlled supervision, so a session that finishes or blocks can ping you where you actually are. Both are documented in the repo's docs section rather than fully in the README, which is the correct place for fast-moving features — treat my summary here as a pointer, not the authority.
Where it falls short
The project says the quiet part loudly: it is actively hunting co-maintainers (pinned issue #1650), with a validation pipeline that triages every PR within about a day. Healthy transparency, but also a signal that one maintainer carries the weight. Speed cuts the other way too — a tool shipping weekly changes muscle memory occasionally, as the v1.9.55 dialog change demonstrates.
One platform note: the list stops at WSL for Windows users; native PowerShell workflows are out of scope. Consistent with house rules, I have not run a nine-session fleet on this — everything above is documented behavior from the repo, pulled September 6, 2026, not personal war stories.
Try it in five minutes
Install, then run the smallest realistic loop:
Attach with Enter, detach with Ctrl+Q, add a second directory from n, and glance at $ to see the cost accounting wake up. If the deck metaphor fits how you already work — multiple agents, multiple projects, one terminal — the five-minute demo tells you immediately.
TL;DR
- Agent Deck is an MIT-licensed Go TUI managing Claude, Gemini, OpenCode, Codex and more in one deck — 845 stars, macOS/Linux/WSL.
- Core commands: agent-deck add . -c claude, session fork, session send --message-file, mcp attach, skill attach, web.
- The keys that matter: Enter/Ctrl+Q to attach/detach, n new, f fork, $ cost dashboard, M groups; Ctrl+S creates sessions since v1.9.55.
- Caveats: WSL-only on Windows, weekly releases move the UI, one maintainer openly recruiting help.
- Five-minute demo: register two directories, attach, detach, check the cost dashboard.
---
Not affiliated with asheshgoplani, Anthropic, or OpenAI. Sources: the agent-deck GitHub repository, its README and release notes, pulled September 6, 2026.