🪴 Brain

herdr

4 min read

herdr

🗒️ herdrdev/herdr — “the runtime your coding agents live on.” A background server that owns the terminals your agents run in, rather than wrapping or replacing the agents themselves. One Rust binary, no Electron. Apache 2.0.

🚀 The problem it fixes: long-running agents die with your terminal, and when you run several you spend your time hunting for the one that stopped and is waiting on an answer.

Description

🧩 What it does:

  • Always running — herdr is a background server and the terminals live inside it. Close the lid, drop the network, restart the machine: agents keep working and sessions come back. Reattach from any terminal, or over SSH.
  • Never hunt for the stuck one — every pane is marked working, blocked or idle. When an agent stops and needs an answer, herdr says so.
  • Agent-native — agents drive herdr through the CLI and socket API: they can spawn panes, prompt each other, and wait until another agent is genuinely blocked. There’s a dedicated agent skill.
  • Runs what you already run — Claude Code, Codex, Cursor, OpenCode, Grok and the rest. It doesn’t wrap them; it owns their terminals.
  • Keyboard and mouse both first-class — tmux-style prefix keys and click, drag, split. Pick per moment, not per tool.
  • Plugins — extend panes and workflows via a marketplace.

Download or use

curl -fsSL https://herdr.dev/install.sh | sh
# or: brew install herdr · mise use -g herdr

Windows:

powershell -ExecutionPolicy Bypass -c "irm https://herdr.dev/install.ps1 | iex"

There’s a separate path for endpoint-protected Windows — relevant on this machine, where Smart App Control has blocked shims before (see Machine Setup Manifest).

Then start it where the work lives:

herdr          # run agents, split panes, walk away
# ctrl+b q detaches; `herdr` reattaches

Reasoning for

This is the session-persistence layer for multi-agent work, and it’s the piece I keep improvising with terminal tabs. Two properties earn it a note:

  1. Survives the machine. A long compile-review-fix loop doesn’t die because I closed the laptop or the VPN dropped. That’s the difference between “agents running overnight” being a real workflow and a fantasy.
  2. blocked is a first-class state. The expensive failure in parallel agent work isn’t a crash — it’s an agent that quietly finished 40 minutes ago and has been waiting for a yes/no. Surfacing that per pane is the actual product.

The agent-native angle matters too: agents can spawn their own panes and wait until another agent is genuinely blocked, which is a primitive the orchestration patterns in Swarm Research — Orchestrating Coding Agents currently have to fake.

Sits at the same layer as the harness questions in Harness Engineering — but below the harness: herdr doesn’t care which agent you run, only that its terminal outlives your session.

Alternatives considered

  • tmux / screen — the honest baseline, and herdr borrows the prefix-key muscle memory. What tmux doesn’t do is know whether the thing inside a pane is working, blocked or idle, or let an agent drive it as a first-class API.
  • Orca — solves an overlapping problem from the GUI end: a desktop/mobile ADE with per-agent git worktrees, diffs and mobile steering. herdr is the terminal-native, headless-friendly, single-binary answer; Orca is the IDE-shaped one.
  • Running agents in IDE terminals — zero setup, but nothing survives a restart and nothing tells you which one is waiting.

Resources

Template: tool

Graph View

Backlinks