Work-oriented AI-agent observability & orchestration for tmux.

See which agents are working, waiting, idle, or blocked from your tmux status line, a live TUI, desktop notifications, and local reports.

English · 한국어

---

Try Muxa onboarding before reading

Launch the complete fullscreen tour—no Muxa binary download or installation:

curl -fsSL https://raw.githubusercontent.com/Open330/muxa/main/scripts/onboard.sh | sh

_{muxa watch — the fleet, the inspector, the swarm view, and muxa attend.}

muxa is a small daemon and CLI for observing — and now driving — AI coding agents running inside terminal multiplexer panes. It reads agent state from existing hook/event systems (Claude Code, OpenAI Codex, Google Gemini CLI), falls back to screen-manifest detection for hook-less agents, and correlates it all with multiplexer panes and sessions. Through muxa mcp a coding agent can also orchestrate the others — inspect state, send prompts, wait for changes.

It does not fork the multiplexer or modify agent binaries. tmux, rmux, and herdr can be observed at the same time; zellij has a CLI baseline. See the Hosts table below.

Optimized for work-oriented tmux

Muxa treats tmux as a durable work execution model, not just a collection of terminal panes:

tmux object	Muxa meaning	How it is used
session	One workspace or project	Durable project context containing several independent work windows.
window	One work item or ticket	Stable work identity and cwd. Starting the same work again reuses this window.
pane	One agent	An implementer, reviewer, or helper working inside that work window.

The intended workflow is equally direct:

1. Start a work ID once; Muxa creates or reuses the workspace session, creates its work window, and starts the first agent pane.
2. Add implementer, reviewer, or helper agents as additional panes in that same work window. Other tickets become sibling windows.
3. Observe, preview, message, and control those agents through muxa watch, or let an agent use the same policy through muxa mcp.
4. Close an agent pane, work window, or whole workspace session explicitly. Muxa refuses to terminate unmanaged tmux objects.

In short: workspace/session → work/window → agent/pane → observe and collaborate → explicit close. Run muxa onboard for one continuous safe scenario. It welcomes you with the reason for creating a practice tmux session, then lets you type tmux new-session, learn the hierarchy, windows, panes, detach/attach, and managed prefix bindings before continuing directly into the Muxa watch workflow. Nothing in the tour mutates a live tmux session.

Important

Beta. Event ingest, the daemon, CLI, live TUI, desktop notifications, stats, and reports work end-to-end, but APIs may still change before 1.0.

What You Get

Surface	What it does
muxa status-line	One-line tmux status-right summary for the active pane.
muxa peek	prefix + q overlay: each pane's live screen dimmed under a box with its agent's state, summary, and latest prompt/response — including how long ago you last prompted it; press a digit to jump.
muxa watch	Main TUI for agents, prompts, live previews, and same-window collaboration.
muxa dashboard	Workspace-card TUI console for pane operations and authenticated same-work agent collaboration.
muxa attend	Jump to the agent blocked on input/choice/error longest.
muxa stats / muxa report	Local analytics for prompt history, agent state duration, tmux foreground time, and human thinking time.
muxa timeline	Full-screen TUI timeline of agent work, waiting, errors, human interaction, and tmux foreground time.
muxa activity	Raw duration ledger query for debugging exactly what fed stats/report.
BarShelf widget (macOS)	Menu-bar popover summary of active, working, waiting, and error agents.
Dashboard	Optional loopback HTTP UI with SSE live updates and a timeline graph.
Notifications	Optional desktop alerts when agents need attention.

Install Muxa

If you decide to keep Muxa, install it with one of the following methods.

Requires tmux 3.x (or herdr) and a Unix-like OS.

Homebrew (pre-built binaries, no Rust toolchain needed):

brew install open330/tap/muxa
muxa init

Or the one-shot installer (builds from source, requires Rust 1.88+):

curl -fsSL https://raw.githubusercontent.com/Open330/muxa/main/scripts/install.sh | sh

Or from source:

git clone https://github.com/Open330/muxa.git
cd muxa
cargo install --path crates/muxad --locked
cargo install --path crates/muxa-cli --locked
muxa init

Verify:

muxad &
muxa status
muxa watch

Collaborate from muxa watch

_{b mailbox · m message the row under the cursor · a/A headless ask.}

The model is simple: one tmux window is one room. The agent pane that is focused when muxa watch opens is the sender.

One-time setup: add the following to ~/.config/muxa/config.toml, restart muxad, and run muxa init to install the prefix+s watch popup.

[collaboration]
enabled = true
wake = "idle_only"

Register the MCP server once for both agent hosts, then restart agents that were already running so they can read and reply to requests themselves.

claude mcp add --scope user muxa -- muxa mcp
codex mcp add muxa -- muxa mcp

Codex only forwards explicitly allowed environment variables to stdio MCP servers. Add this line to the generated [mcp_servers.muxa] table in ~/.codex/config.toml (especially when using a custom muxa/tmux socket):

env_vars = ["RMUX", "RMUX_PANE", "TMUX", "TMUX_PANE", "MUXA_SOCKET"]

Muxa also recovers the pane from process ancestry across active pane backends for existing default-endpoint Codex registrations, so older setups fail safely rather than appearing paneless.

Connected agents are told that room peers can serve as read-only reviewers or narrowly scoped execution subagents. Requests and replies can also carry validated AIR 1.0 artifact references, which watch visualizes with profile-colored mailbox badges.

Agents reported as synthetic by muxa doctor are omitted from collaboration until a hook event establishes a stable session identity. Submit a prompt or restart that agent, then check again.

Then:

1. Run two agents in two panes of the same tmux window.
2. Focus the agent that should send, then press prefix+s.
3. Select the other agent in watch, press m, type the request, and press Enter. Press M to read and reply from the mailbox (b remains an alias).

Do not open watch from a spare shell pane when you want to collaborate: that shell is not an agent. Continue the collaboration workflow in watch. For request/reply details, see docs/COLLABORATION.md.

For install modes, muxa init presets, systemd, manual hook wiring, and rollback details, see docs/INSTALL.md.

Core Commands

Managed tmux policy: one session is a workspace/project, each window is a work/ticket, and each pane is an agent. muxa onboard teaches tmux first and introduces this Muxa mapping only after the tmux exercises, as one continuous scenario. It starts at a blank virtual shell, accepts the real tmux new-session -s muxa-onboarding and tmux attach -t muxa-onboarding commands, and explains why each session command is needed before asking for it. It preserves every virtual window/pane transition, then continues without leaving fullscreen into the current muxa watch workflow. The watch half mirrors the left-edge session-state gutter, columns, 50/50 inspector, overlays, and footer. Commands and keys that you must enter are shown in bold yellow in both the dialog body and footer. You advance with the real j, l, Alt-T, o, ?, n, m, Backspace, M, and q actions. One 20-step counter covers the whole scenario: managed prefix bindings are step 11 and work navigation follows as step 12. Korean is selected automatically for a Korean locale, can be requested with --lang ko, and can be toggled with F2 during the tour.

_{muxa onboard — shell, tmux, and Muxa in one safe interactive scenario.}

Command	Purpose
muxa status [--json]	Human-readable table, or a versioned JSON snapshot for desktop integrations.
muxa watch [--view pane|work]	Live workspace → work → agent TUI picker/dashboard.
muxa dashboard [--since today]	Workspace-card TUI with live capture, prompt/actions, ACT/WACT totals, and collaboration mailbox controls.
muxa attend [--cycle] [--list]	Focus or list agents needing attention.
muxa status-line [--pane %N]	tmux status-line output.
muxa peek [--plain]	Per-pane overlay for the current tmux window; --plain prints it as text.
muxa recap [--pane %N]	Recent prompts from retained disk history.
muxa peers / muxa identity / muxa msg	Discover and name same-window agents, then exchange durable request/reply messages.
muxa stats --since today	Focused WACT/ACT/WORK/WAIT summary; group by day/project/agent/session. Add --graph for graph-only WACT over time or --verbose for diagnostic columns.
muxa report --since week	All breakdowns (day/project/agent/session) as focused ACT/WACT tables; add --json or --markdown to export.
muxa timeline --since today	Interactive session-grouped timeline; filter with --session main / --agent codex, sort with --sort waiting, or use --view heatmap.
muxa activity --type agent|tmux|human	Raw activity ledger intervals.
muxa sync	Backfill the registry by scanning active pane hosts.
muxa register --name X [--pid N]	Surface an arbitrary background process (script, game, automation loop) as a pid-tracked row in muxa status.
muxa run --detach --name X -- <cmd>	Run a command in a muxa-owned PTY; it also appears in muxa status as a task.
muxa work start muxa-onboarding --workspace muxa --agent codex ...	Create/reuse workspace session muxa, create/reuse its work window, and add an agent pane.
muxa workspace list/show/close	Inspect or explicitly close workspace/project sessions.
muxa work list/show/close [--workspace muxa]	Inspect or explicitly close work/ticket windows.
muxa agent start --workspace muxa --work muxa-onboarding ...	Add an allowlisted agent pane to one work window; also exposed as MCP muxa_start_agent.
muxa agent control --pane %N --action interrupt	Interrupt or explicitly terminate one managed agent pane.
muxa onboard [--lang auto|en|ko]	Unified shell → tmux → Muxa fullscreen walkthrough. F2 switches language, --no-quiz skips gates, and --print emits the combined guide.
muxa mcp	MCP stdio server so a coding agent can orchestrate muxa — inspect agents, send prompts, capture panes, wait for changes (claude mcp add --scope user muxa -- muxa mcp, see docs/MCP.md).
muxa init	Interactive install/uninstall wizard.
muxad	Daemon process.

Common stats queries:

muxa stats --since today --group-by session
muxa stats --since yesterday --group-by project
muxa report --since last-week
muxa timeline --since today --session main
muxa timeline --since today --exclude-session 'monitor*'
muxa stats --since month --exclude-pane '%42' --exclude-session 'monitor*'
muxa timeline --since today --group-by kind --sort waiting
muxa timeline --view heatmap --since 12w
muxa timeline --day 2026-06-06
muxa activity --since today --type human

--since accepts today, yesterday, week for a rolling 7-day window, month for a rolling 30-day window, last-week / "last week" for the previous Monday-Sunday calendar week, last-month / "last month" for the previous calendar month, rolling durations like 24h/7d/4w, local dates like 2026-06-06, RFC3339 timestamps, and all. See docs/ACTIVITY.md for ledger semantics, including HUMAN, THINK, and ACT.

muxa stats, muxa report, and muxa timeline also accept --exclude-pane and --exclude-session for long-lived monitoring scopes. Patterns are case-sensitive and support * and ?, e.g. --exclude-session 'monitor*'.

Supported Agents

Hook-based (authoritative). These wire into their existing hook/event systems, so muxa gets exact state transitions:

Agent	Status	Config
Claude Code	Supported	~/.claude/settings.json
OpenAI Codex	Supported	~/.codex/config.toml
Google Gemini CLI	Supported	~/.gemini/settings.json
opencode	Planned	tracking issue

Screen-detected (fallback). Agents with no hooks are classified from their pane contents via TOML manifests — bundled best-effort for cursor-agent, amp, copilot, aider, and goose, extensible per user. Hooks always win when present. See docs/SCREEN_DETECTION.md.

On herdr hosts, muxa additionally surfaces every agent herdr's own detection sees, with no manifest needed.

Hosts

muxa observes agents across terminal-multiplexer backends and can watch several at once (e.g. during a tmux→herdr migration):

Host	Status	Notes
tmux	Full	The default backend.
rmux	Initial CLI backend	Pane discovery, capture, focus, and targeted input; see docs/RMUX.md.
herdr	Full	Via herdr's socket API; see docs/HERDR.md.
zellij	CLI baseline	Richer plugin path planned; see docs/ZELLIJ.md.

See docs/MULTI_HOST.md for observing multiple hosts simultaneously.

More Docs

Topic	Doc
Install and wiring	docs/INSTALL.md
Onboarding and work/agent policy (한국어)	docs/ONBOARDING.ko.md
MCP control plane (muxa mcp)	docs/MCP.md
herdr host support	docs/HERDR.md
rmux host support	docs/RMUX.md
Multi-host observation	docs/MULTI_HOST.md
Screen-manifest detection	docs/SCREEN_DETECTION.md
Live TUI and prompt composer	docs/WATCH.md
CLI dashboard	docs/DASHBOARD_CLI.md
Stats, reports, activity ledger	docs/ACTIVITY.md
Timeline TUI and dashboard graph	docs/TIMELINE.md
Configuration reference	docs/CONFIGURATION.md
Web dashboard	docs/DASHBOARD.md
External sinks	docs/SINKS.md
Zellij plan	docs/ZELLIJ.md
Architecture and development	docs/ARCHITECTURE.md
Agent collaboration	docs/COLLABORATION.md

Development

cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all -- --check

License

MIT OR Apache-2.0.