operad

Cross-platform tmux session orchestrator for Claude Code sessions. Designed for developers running multiple Claude Code projects simultaneously.

What it does:

• Boot and manage tmux sessions with dependency ordering
• Health checks, auto-restart, and session lifecycle management
• Web dashboard: session status, memory, logs, telemetry, settings
• Prompt history: search, star, and replay Claude prompts across all projects
• Battery and memory awareness on Android/Termux
• Token quota tracking with velocity trends and per-session attribution

Quick Start

npm install -g operadic

mkdir -p ~/.config/operad

# Create config with your projects
cat > ~/.config/operad/operad.toml << 'EOF'
[operad]
dashboard_port = 18970

[[session]]
name = "my-project"
type = "claude"
path = "$HOME/git/my-project"
EOF

operad boot
# Dashboard: http://localhost:18970

Run operad doctor to diagnose any setup issues.

Config

Default location: ~/.config/operad/operad.toml (TOML with $ENV_VAR expansion)

[operad]
dashboard_port = 18970
health_interval_s = 120
wake_lock_policy = "active_sessions"

[[session]]
name = "my-project"
type = "claude"
path = "$HOME/git/my-project"
auto_go = true

[[session]]
name = "api-server"
type = "service"
command = "bun run dev"
path = "$HOME/git/api"
depends_on = ["my-project"]

[session.health]
check = "http"
url = "http://localhost:3000/health"

Session types: claude (Claude Code with readiness detection), daemon (long-running command), service (headless).

CLI Commands

| Command | Description | |---------|-------------| | operad boot | Start daemon + all sessions in dependency order | | operad status | Session table with memory, battery, uptime | | operad health | Run health sweep | | operad start <name> | Start a specific session | | operad stop <name> | Stop a specific session | | operad restart <name> | Restart a specific session | | operad go <name> | Send "go" to a Claude session | | operad open <path> | Register and start a dynamic session | | operad close <name> | Stop and unregister a dynamic session | | operad recent | Recent Claude projects from history | | operad tabs | Open terminal tabs for running sessions | | operad memory | System memory + per-session RSS | | operad suspend <name> | SIGSTOP a session | | operad resume <name> | SIGCONT a session | | operad logs | Stream daemon logs | | operad doctor | Diagnose install issues | | operad upgrade | Rebuild and hot-swap daemon | | operad shutdown | Stop daemon (sessions persist in tmux) |

Dashboard

The web dashboard at http://localhost:18970 provides:

• Overview — session status, system memory, budget gauges, prompt history
• Memory — per-session RSS tracking, AI memory management (SQLite + FTS5), process manager
• Logs — real-time daemon logs with level filtering
• Telemetry — captured telemetry sink with SDK breakdown
• Settings — MCP servers, plugins, skills, plans, CLAUDE.md management

Platforms

| Feature | Android/Termux | Linux | macOS | Windows (experimental) | |---------|---------------|-------|-------|------------------------| | Notifications | termux-notification | notify-send | osascript | PowerShell toast | | Battery | termux-battery-status | /sys/power_supply | pmset | WMI (laptop only) | | Wake lock | termux-wake-lock | systemd-inhibit | caffeinate | not implemented | | Process info | /proc | /proc | ps/lsof | tasklist (limited) | | Terminal tabs | am intents | n/a | Terminal.app | n/a | | ADB protections | phantom fix + Doze | n/a | n/a | n/a |

Windows requires tmux via MSYS2 (pacman -S tmux) or WSL. See docs/windows.md for setup.

Crash Resilience

On Android, the daemon, watchdog, and tmux server all run as independent processes (PPid: 1). When Android kills the Termux app, only the terminal UI dies — all sessions continue running. The watchdog auto-restarts the daemon, which re-adopts existing sessions.

Defense layers: wake lock (never released), phantom process killer fix, Doze whitelist, process detach, IPC socket self-healing, watchdog loop, crash-safe trace log.

Advanced: Autonomous Layer

Opt-in. Disabled by default. These features run AI agents autonomously. Enable via dashboard Settings → Switchboard after reading the in-app docs.

operad includes an agentic layer for self-improving orchestration:

• OODA loop — periodic Observe→Orient→Decide→Act cycles via master-controller agent
• Agents — optimizer, preference-learner, ideator, master-controller
• Scheduling engine — cron/interval triggers for agents and commands
• Memory system — decay, consolidation, cross-pollination of agent learnings
• Agent specialization — domain expertise tracking with roundtable protocol for multi-agent collaboration
• Tool registry — extensible tool system with autonomy levels, trust calibration, and persistent leases
• Tuning — feed notes, personality traits, and chat logs to shape autonomous decisions

See in-app /help for full documentation.

Development

bun install
bun run build       # bundle to dist/tmx.js
bun run typecheck   # TypeScript check
bun test            # unit tests
cd dashboard && bun run build  # build dashboard

Docs

operad.stream

License

MIT