Agent Deck — Summary

Agent Deck is a Go + Bubble Tea terminal session manager designed as mission control for fleets of AI coding agents. It renders a full-screen TUI that lets users create, start, stop, fork, and switch between multiple Claude Code or OpenCode sessions with single-keystroke speed. The conductor subsystem lets one long-running "head" session supervise workers, relay Telegram/Slack notifications, and dispatch sub-agents for independent tasks. An MCP manager and Skills manager allow toggling per-session MCP servers and Claude skills without editing JSON configs. State persistence (state.json + task-log.md + LEARNINGS.md + HANDOFF.md) survives Claude Code context compaction and process restarts.

Agent Deck occupies a unique niche as pure infrastructure middleware: it does not define prompts, phases, or methodology — it only manages the process layer around existing AI coding agents. Compared to claude-flow (the nearest seed), Agent Deck provides the same "fleet" framing but achieves it with a thin Go binary + tmux-pane model rather than an npm package + MCP server + 305 MCP tools. Agent Deck's UI surface (Bubble Tea TUI + optional web UI) is a first-class engineering primitive, while claude-flow has no local UI at all.

Agent Deck — Overview

Origin

Agent Deck was created by Ashesh Goplani and first released in 2025. It grew from the need to manage parallel Claude Code sessions across many projects without hunting through terminal tabs. The repo has 30 contributors and 2,535 stars as of 2026-05-26.

Philosophy

The project positions itself as "mission control for your AI coding agents." The central thesis: running multiple AI agents concurrently is already possible with Claude Code, but the UX for managing them is broken (tab proliferation, no visibility into which agents are running/waiting/stuck). Agent Deck solves the visibility and control problem without changing anything about how the agents themselves work.

It deliberately avoids methodology lock-in. There are no SPARC phases, no SDD gates, no persona files. The value is purely operational: see all agents, switch fast, attach tools dynamically.

Key manifesto points from the README

• "One terminal. All your agents. Complete visibility."
• "Running Claude Code on 10 projects? OpenCode on 5 more? Another agent somewhere in the background? Managing multiple AI sessions gets messy fast."
• "Fork any Claude conversation instantly. Each fork inherits the full conversation history."
• "Attach MCP servers without touching config files."

Conductor pattern

The conductor is Agent Deck's highest-level abstraction: a long-running Claude session configured to receive heartbeats, answer Telegram/Slack messages, escalate to the user, and dispatch sub-agent workers. The README frames this as "from now on every other agent-deck session you run is supervised by a single conductor session that answers routine questions, escalates the interesting ones to your phone, and never lets a waiting worker rot."

Agent Deck — Architecture

Distribution

• Binary: Go-compiled single binary agent-deck, cross-platform (macOS, Linux, WSL)
• Install: install.sh curl-pipe or pre-built releases from GitHub
• Claude Code plugin: "/plugin marketplace add asheshgoplani/agent-deck" → installs SKILL.md + references

Required Runtime

• Go 1.24+ (to build from source)
• Pre-built binary: no runtime dependency
• tmux (for pane management behind the TUI)
• Claude Code or OpenCode (the agents being managed)

Directory Tree

agent-deck/
├── cmd/              # Go CLI entry point
├── conductor/        # Conductor orchestration logic
├── internal/         # Core session/MCP/profile logic
├── skills/
│   └── agent-deck/
│       ├── SKILL.md                    # Claude Code skill
│       └── references/
│           ├── cli-reference.md
│           ├── config-reference.md
│           ├── tui-reference.md
│           └── troubleshooting.md
├── scripts/          # Helper scripts (launch-subagent.sh, etc.)
├── site/             # Static documentation site
├── sandbox/          # Test sandbox environment
├── .claude-plugin/
│   └── marketplace.json
└── ~/.agent-deck/    # User config (runtime, not in repo)
    └── config.toml

Config Files

• ~/.agent-deck/config.toml — profiles, groups, conductors, MCP pool, per-group Claude config override
• .agent-deck/skills.toml — per-project skills state (applied vs pool)
• state.json, task-log.md, LEARNINGS.md, HANDOFF.md — per-session state persistence (workspace directory)

Target AI Tools

Claude Code, OpenCode (auto-discovered from PATH)

Tech Stack

• Core: Go 1.24, Bubble Tea (charmbracelet/bubbletea), Lip Gloss (charmbracelet/lipgloss)
• State: SQLite (modernc.org/sqlite)
• Notifications: gorilla/websocket, Google Cloud Pub/Sub (optional)
• Web UI: served from within the binary (esbuild-bundled React)

Agent Deck — Components

CLI Subcommands (binary: agent-deck)

TUI Surfaces

Skills (Claude Code plugin)

Scripts (within skill)

State Files

Agent Deck — Uniqueness & Positioning

differs_from_seeds

Agent Deck is closest to claude-flow in goal (fleet orchestration of multiple AI coding agents) but diverges architecturally in every dimension. Claude-flow installs an MCP server with 305 tools, relies on npm/npx, stores state in SQLite+vector, and has no UI surface. Agent Deck installs a single Go binary with a Bubble Tea TUI, manages sessions via tmux panes, and stores state in SQLite + markdown files. Claude-flow defines SPARC methodology phases and generates spec files; Agent Deck defines no methodology at all. Claude-flow is an MCP-anchored toolserver (Archetype 3); Agent Deck is a pure process-management layer with no prompt methodology (no archetype match in the seed taxonomy — closest to Archetype 4 "markdown scaffold, zero primitives" but in reverse: all CLI/TUI, almost zero prompt content). The conductor+worker pattern overlaps superficially with claude-flow's queen+worker pattern, but Agent Deck's conductor is a plain Claude Code session given a description file, not a specialized code-class subagent.

Positioning

• Compared to claude-conductor (seed): claude-conductor is 15 markdown templates with no runtime. Agent Deck is the opposite: no templates, all runtime.
• Compared to superpowers: superpowers injects methodology via skills; Agent Deck provides the operational harness for whichever methodology already exists.
• First-class GUI multi-session management is absent from all 11 seeds. Agent Deck fills that gap for terminal users.

Observable Failure Modes

1. tmux dependency: If tmux is not installed or the system limits panes, sessions silently fail.
2. Conductor single-point-of-failure: The Telegram relay depends on the conductor session staying alive; if it compacts or crashes, worker escalations queue up undelivered.
3. Script path resolution: The SKILL.md dedicates a full "IMPORTANT" section to preventing agents from using the wrong script path — this is a known recurring failure mode.
4. Partial Web UI: The web dashboard is documented as partially functional; users relying on it for MCP management may hit unverified features.
5. No methodology enforcement: Agent Deck provides no quality gates, TDD enforcement, or review steps. All agent discipline must come from the agents themselves or separate skill packs.