herdr

• herdr

  • Terminal-native agent multiplexer: surfaces the state of every coding agent at a glance — blocked, working, done — as real terminal panes rather than a wrapped interpretation of them. Self-described as "the runtime your coding agents live on."
  • herdrdev/herdr — GitHub/Star: 24,249 (checked 2026-08-04). Apache 2.0.
  • Written in Rust; ships as a single binary rather than an Electron app, and runs inside whatever terminal is already in use.

  • Where it sits against tmux

    • Covers the multiplexer basics that make it a plausible tmux replacement rather than a companion: detach and reattach from any terminal, sessions that survive restarts, and attach over SSH. Weighed on Terminal/Multiplexer/Idea.
    • Keeps tmux-style prefix keys while treating mouse click, drag, and split as equally first-class, so neither input mode is the fallback.

  • What the agent orientation adds

    • Agent state detection is the premise: panes are labeled by whether the agent inside is blocked, working, or done, which is the piece a general-purpose multiplexer leaves to the human to infer.
    • A socket API lets agents drive the multiplexer themselves — spawning panes, reading output, and waiting on each other — rather than only being driven by a human at the prefix key.
    • Plugins extend panes and workflows, distributed through a marketplace.

  • Adoption

    • Person/Chris Power switched over after about 10 years on tmux, won over by the agent-state pane and by being able to SSH into a remote box and leave an agent running unattended — Person/Chris Power/YouTube/26/08/herdr.
    • Installed from My/Dotfiles since 2026/08/11 as a global mise tool, alongside tmux rather than replacing it, to try the agent-state pane in daily use.
      • Config lives at ~/.config/herdr/config.toml and holds two settings. onboarding = false stops herdr's first-run write from editing a chezmoi-managed file; [update] version_check = false leaves the version to mise, which herdr already defers to. Its defaults supply the rest — the ctrl+b prefix and Catppuccin/Mocha.
      • Detection manifests stay on: they refresh the screen-pattern rules behind the blocked/working/idle labels, which decay as agent interfaces change.

  • No longer the primary multiplexer here, since 2026/08/24

    • Person/codekiln on 2026-08-25 Tue: "herdr is not retired; it's just not primary as my multiplexer right now." The software is alive and this graph said otherwise for a day.
    • The server was stopped on the morning of 2026/08/24 and its ~24 orphaned sessions killed. My/AI/Agent/Fleet runs on tmux now, and the standing instruction is not to start the herdr server or use herdr commands. That instruction is about which multiplexer the seats run, and it carries no verdict on the tool.
    • The reason was hierarchy rather than any fault in the agent-state pane. A herdr workspace holds panes, and the agent name carried the whole address in 32 characters, so <workspace-label>-<tab-label> had to encode the repository and the work item together. tmux names five nested levels separately — session, window, pane, claude session, branch and worktree — and C-b w walks all of them in one picker.
    • What went with it is the blocked/working/done label read off the pane border, which tmux has no equivalent for. The fleet reads a session's state from its own transcript instead — see My/AI/Agent/Fleet/Bed Down on ctx-check.
    • Orphaned sessions were the practical cost of the run: nothing reaped a workspace whose agents had finished, so they accumulated until the server was stopped by hand.