A tmux-based bullpen for your AI coding agents

What it does

CCB is a project-level workspace that runs multiple real coding-agent CLIs — Codex, Claude, Gemini, OpenCode, Droid, Antigravity — as supervised tmux panes. A TOML config (.ccb/ccb.config) defines the team: which agents exist, which provider each uses, how they’re grouped into windows, and who gets their own git worktree. Agents delegate to each other through explicit /ask and callback routes, and a sidebar shows who’s doing what.

The interesting bit

The honest angle: this is orchestration glue, and it knows it. Rather than hiding agents behind an abstraction layer, every pane is the actual provider CLI — you can click into it, type, scroll its history, and see exactly what it’s doing. The README is also unusually candid about positioning, with a comparison table against Claude Code subagents and Hive instead of pretending competitors don’t exist.

Key highlights

• Mixed providers in one project, with per-agent model, key, and API settings
• Per-agent git worktrees so parallel workers don’t trample each other
• Recoverable runtime: panes are supervised, with attach, restore, and project-scoped cleanup
• Config reload with a dry-run preview, without restarting other agents
• Runs on Linux, macOS, and WSL; Python 3.10+

Caveats

• Not an automatic DAG generator — you design the topology explicitly in config, which is the whole philosophy but also real work
• Requires tmux fluency (or at least tolerance); the README devotes a long section to it for a reason
• Install is release-first; source installs are explicitly relegated to development/fallback use

Verdict

Worth a look if you already run multiple agent CLIs and want them coordinated in the open instead of behind a harness. If you want automatic workflow graphs or production-grade state and recovery, the README itself points you elsewhere.