Claude Tmux Orchestration

Claude Code tmux Orchestration - Parallel AI Sessions with Full Context
Spawn full Claude Code sessions as workers, coordinated through a bash heartbeat and file-based state. Attach to any of them mid-run and take over by hand.

Status: read this before you clone

Built March 2026. Reviewed 2026-07-31 against current Claude Code. Still runs, but the reason to use it has changed.

This was built when subagents could not do much. Since then Claude Code absorbed most of it. Per Anthropic's subagent docs, subagent frontmatter now accepts maxTurns, hooks, mcpServers, skills, memory, effort, background and isolation, and each subagent runs in its own context window.

So the original pitch - "subagents are crippled, use full sessions instead" - is no longer true. Try subagents first. They are one file instead of a bash control loop.

What this still buys you

Three things survive, and only three. If none of them matter, use the native tools.

- A session you can attach to. tmux attach, watch it work, interrupt it, answer its question by hand, let it continue. No lost context. When a long autonomous run goes sideways at minute 40, taking the wheel is worth the whole control loop.
- Full session lifecycle. Compaction, /clear, /resume, slash commands, mid-run model switching via send-keys.
- An orchestration loop you own. Bash and JSON files on disk. Runs on a server, across repos, over SSH, on a schedule. No dependency on a product feature staying put.

Also worth knowing: Agent Teams is the first-party option, still behind CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, and its split-pane display mode needs tmux or iTerm2 anyway. Dynamic workflows orchestrate subagents at scale from a script, up to 16 concurrent and 1,000 per run.

Architecture Overview

The 4-Step Monitoring…