Herdr

Operator's take

The failure mode of desktop agent apps is quiet and total: quit the window, close the laptop, or crash the app, and the herd dies with it. Herdr's answer is architectural rather than cosmetic — the thing that owns your agents is a server, and every interface is just a client of it. The TUI in the screenshots can attach, detach, and even crash without the agents noticing; so can the CLI, plain SSH, and whatever client comes next.

What that looks like day to day: a terminal multiplexer in the tmux lineage — real processes in real PTYs, detach and reattach, sessions that survive the terminal closing — that also understands agents. It knows which pane is an agent and whether it's blocked, working, done, or idle, and it knows how to wait on an agent instead of polling it. The CLI and a JSON socket API expose all of it, which is the part agents themselves can drive: read output, send input, wait on status, split panes. Your terminal emulator stays (Ghostty, Kitty, iTerm, Alacritty), there's no web view, no account, and no hosted control plane.

The catch is that Herdr is deliberately the layer under, not the app on top. No worktree management, no diff review, no task queues — its own comparison page says it "pairs happily with a worktree manager" rather than being one. If you want one window that owns the whole build-and-review loop, this is the wrong tool; if you want agents that survive your UI and are reachable from a server, a Mac mini, or a phone over SSH, this is the most direct bet in the lane.

What it's good at

• Agents outlive the UI — a server owns the terminals; TUI, CLI, and plain SSH clients attach and detach freely while work keeps running.
• Semantic agent state — blocked, working, done, and idle at a glance across the whole session, not just pane titles.
• Anywhere access — runs wherever the work is (a server, a Mac mini, a sandbox VM, anything you can SSH to), with a switcher built for phone-sized screens when you attach from mobile.
• Native remote attach — a thin-client mode (herdr --remote workbox) that installs Herdr on the host for you, bridges your clipboard including image paste, and keeps your keybindings.
• An API agents drive themselves — CLI plus a JSON socket: create workspaces and tabs, split panes, run work, wait on agent status, read output.
• Plugin ecosystem — community plugins auto-discovered from any public GitHub repo tagged herdr-plugin, installable in one command; 986 listed at capture.
• Wide agent support — 21 agent CLIs detected, from Claude Code, Codex, and Cursor to Pi, Droid, Amp, OpenCode, Grok CLI, Kimi, Kiro, and Copilot; integrations add richer state and session resume.

What it's not

• Not a manager app — no worktrees, no diff review, no review queues; it pairs with those tools instead of replacing them.
• Terminal-native by conviction — no web dashboard and no desktop GUI is a design decision, not a gap; if the terminal isn't where you want to live, this is the wrong bet.
• Self-run runtime — you run the server, which is the point and also the ops burden: updates, uptime, and the box it sits on are yours.
• Young and fast-moving — large install base and daily change; expect surface churn.