herdr Ships a Desktop App and herdr-core: a UI Your Agents Can Rewrite

Running three or four Claude Code and Codex sessions at once, flipping between windows to find the one stuck waiting on you, has become a daily routine for developers this year. The open-source project herdr's answer is to give those agents a proper home.

herdr started as a terminal app, now called Herdr TUI: a single Rust binary, no Electron, with a background server that owns every terminal so agents keep working when you detach or close the laptop. Now the team has shipped a brand-new native desktop app plus the local daemon behind it, herdr-core. All three are open source under Apache 2.0.

The window is a skin; the agents live in core

herdr-core is a small server on your machine where every agent lives. Quit the app, update it, crash it, and every terminal is still where you left it. The design cleanly separates interface from runtime: the window is just a client you can close or swap at will.

It isn't picky about agents either. Claude Code, Codex, Pi, OpenCode, Gemini, Copilot, Cursor, Amp: anything that runs in a terminal works. herdr reads each agent's status (working, done, needs you) through an open protocol, OSC 7501, the Program Status Protocol. One printf line turns your own script into an agent too.

Multiple machines are covered. Your laptop, the desktop upstairs, a rented VPS, a cloud sandbox: anything reachable over SSH joins the same sidebar. The laptop sleeps, the remote agents keep going.

"Every pixel is a file"

The most interesting idea in the desktop app is its tagline: made for your agents, remade by them. Everything you see is a plain file in ~/.config/herdr-app, with an AGENTS.md telling your agent what it may change.

So there is no plugin system and no fork to maintain. Want a different sidebar or a new button? Ask your agent to edit it. It is a very 2026 view of software: rather than design an extension API, ship the product in a form agents can read and write directly. That won't suit every product, but for a tool built around agents it is entirely coherent.

Agents directing agents

herdr also treats agent collaboration as a first-class feature. One agent can use hc run to start a helper in the background to review a diff, and hc send to redirect it or answer its questions. Reports, exits and "needs you" alerts land straight in the lead agent's input. The human only sees the result.

That reflects how coding-agent usage is changing: from one person watching one agent to one person managing a fleet of agents that manage each other. tmux solved the problem of a person juggling many shells; herdr wants to be tmux for the agent era.

There are limits, and herdr is upfront about them. If the machine reboots or herdr-core stops, the programs stop too; panes come back as shells in the same folder, but agents don't resume work in progress. The desktop app and the TUI are separate programs that don't share agents.

The desktop app runs on Apple-silicon macOS, Linux on x86_64 and arm64, and x86_64 Windows. With Cursor, Anthropic and OpenAI all building their own agent dashboards, a neutral, open-source, model-agnostic home for agents may be exactly what heavy users actually want.