Agent Deck
A terminal TUI command center for managing multiple AI coding agent sessions (Claude, Gemini, Codex, and more) with session forking, MCP management, and conductor orchestration.
At a Glance
Fully free and open-source under the MIT License. All features included.
Engagement
Available On
Listed May 2026
About Agent Deck
Agent Deck is an open-source terminal UI built in Go that acts as mission control for AI coding agents. It runs on macOS, Linux, and Windows (via WSL), and is distributed under the MIT License. The project was created in December 2025 and reached v1.9.24 by May 2026, accumulating over 2,400 GitHub stars according to the repository page.
What It Is
Agent Deck is a tmux-backed session manager purpose-built for AI coding tools. Instead of juggling dozens of terminal tabs, users get a single TUI that shows every running agent — Claude Code, Gemini CLI, OpenCode, Codex, Copilot, and others — with real-time status indicators (running, waiting, idle, error). It adds AI-specific intelligence on top of tmux: smart status detection, session forking with full conversation-history inheritance, MCP server management, and organized session groups.
Core Workflow
The typical workflow centers on the TUI launched by running agent-deck. From there, users can:
- Add sessions with
agent-deck add . -c claudeto attach an AI tool to a directory - Fork sessions with
f/Fto branch a conversation and explore multiple approaches without losing context - Toggle MCP servers per-session or globally via the MCP Manager (
m), with automatic restarts - Search across all sessions with
/(fuzzy) orG(global across all Claude conversations) - Track costs in real time via the
$cost dashboard, which covers 14 models including Claude Opus 4.x, Gemini, GPT-4o, and o3
A web UI (agent-deck web) provides a browser-based view with Chart.js cost charts, SSE live updates, and optional bearer-token protection.
Conductor and Orchestration
The standout feature is the Conductor system — persistent agent sessions that monitor and orchestrate all other sessions. A conductor watches for sessions that need input, auto-responds when confident, and escalates to the user via Telegram or Slack when it cannot. Multiple conductors can run per profile, each with its own identity, Claude config, and channel pairing. The setup wizard (agent-deck conductor setup) handles Telegram/Slack wiring in a single command.
Complementing conductors are Watchers — event adapters (webhook, GitHub, ntfy, Slack) that forward inbound events to conductor sessions so agents can react to external triggers like GitHub PRs, push notifications, or calendar events.
Architecture and Platform Details
Agent Deck is built with Go 1.24+ and uses the Bubble Tea TUI framework. Key architectural choices include:
- tmux socket isolation (v1.7.50+): an opt-in
socket_nameconfig keeps agent-deck's tmux server fully separate from the user's interactive tmux - MCP socket pooling: sharing MCP processes across sessions via Unix sockets, which the README claims reduces MCP memory usage by 85–90%
- Docker sandbox: sessions can run inside isolated containers with bind-mounted project directories; host tool auth is automatically shared into containers
- Git worktree integration: each session can operate in its own worktree and branch, with merge/cleanup commands built in
- SQLite state: session metadata, cost events, and socket assignments are persisted locally
Multi-Tool Support
Agent Deck supports a range of terminal-based AI tools at varying integration depths:
| Tool | Integration Level |
|---|---|
| Claude Code | Full (status, MCP, fork, resume) |
| Gemini CLI | Full (status, MCP, resume) |
| OpenCode | Status detection, organization |
| Codex | Status detection, organization, conductor |
| Copilot | Organization, launch |
| Crush (charmbracelet) | Status detection, organization, launch |
| Custom tools | Configurable via [tools.*] in config.toml |
Update: v1.9.24
The latest release as of May 2026 is v1.9.24, published on 2026-05-20. Recent version history visible in the README includes v1.9.10 (bare-repo worktree support), v1.7.60 (group-scoped Alt+ keybindings), v1.7.58 (nested bare-repo layout), v1.7.50 (tmux socket isolation), v1.7.41 (feedback prompt pacing), v1.7.34 (per-session transition-notify suppression), and v1.5.4 (per-conductor Claude config). The project shows active weekly release cadence and a growing feature surface around multi-agent orchestration, cost tracking, and remote instance management.
Community Discussions
Be the first to start a conversation about Agent Deck
Share your experience with Agent Deck, ask questions, or help others learn from your insights.
Pricing
Open Source
Fully free and open-source under the MIT License. All features included.
- All TUI features
- Conductor orchestration
- MCP manager
- Cost tracking
- Docker sandbox
Capabilities
Key Features
- TUI session manager for multiple AI coding agents
- Real-time status detection (running, waiting, idle, error)
- Session forking with full conversation history inheritance
- MCP server manager with per-session or global toggle
- Skills manager for Claude skills per project
- Conductor system for orchestrating and monitoring agent fleets
- Telegram and Slack integration for remote conductor control
- Watcher adapters (webhook, GitHub, ntfy, Slack) for event-driven orchestration
- Cost tracking dashboard across 14 AI models
- Git worktree integration with merge and cleanup commands
- Docker sandbox for isolated agent sessions
- MCP socket pooling to reduce memory usage
- tmux socket isolation (opt-in)
- Web UI with Chart.js cost charts and SSE live updates
- Remote instance management over SSH
- Fuzzy search and global search across all sessions
- Per-group and per-conductor Claude config overrides
- Budget limits with warning and hard stop
- Notification bar for waiting sessions in tmux status bar
- Multi-tool support: Claude Code, Gemini CLI, OpenCode, Codex, Copilot, and more
