vladzima/herd

Herd is a small coordinator that lets one AI agent manage other AI agents for you. You give a task to a "head" agent running Claude Code, and it delegates pieces of that task out to "worker" agents, each running in its own terminal tab. Workers can be Claude, Codex, Gemini, or other agent tools. The head watches their progress, collects their reports, and summarizes the results back to you. The project is just three files. herd-up starts the head tab. herd-spawn creates a new worker tab with a specific task. A skill file acts as the coordinator's playbook, telling the head how to assign work, check on workers, and handle situations where a worker gets stuck. Because the playbook installs as a Claude Code skill, any Claude Code session can become a head by invoking the /herd command. Communication works through report files. Each worker writes its final report to a file in a temporary directory, and the head treats that file as the real signal that work is done, rather than relying on whether the worker process is still running. If a worker hits a routine question, the head answers it automatically. If something risky comes up, the head escalates to you instead of deciding on its own. To use it, you need herdr (version 0.7.5 or later) and Claude Code installed, plus a tool called jq. Installation is a short shell script that copies the files into local directories. You can run herd-up from inside a project, then type tasks directly to the head or send prompts from elsewhere. You can also spawn workers yourself, choosing the agent type and passing custom arguments, if you want to skip the head entirely.