A coding agent that dies the moment you close the terminal window isn’t an agent — it’s a script with extra steps. Anyone who has handed a long refactoring job to Claude Code or Codex knows the feeling: the train enters a tunnel, the Wi-Fi drops, the SSH connection is gone, and half an hour of work goes with it.
That is the gap Herdr aims to close. The project describes itself as “the runtime your coding agents live on.” Technically it is a terminal multiplexer — but one designed for agents rather than humans from the outset. In September 2026 it raised a six-million-dollar seed round led by Bessemer Venture Partners, which is reason enough to look at what is actually being built.
Table of Contents
Key Takeaways
- Herdr is a terminal multiplexer written in Rust that runs AI coding agents in persistent background sessions, surviving disconnections and machine restarts.
- It reads the output of every pane and automatically marks each agent as working, blocked, or idle. No more clicking through twenty terminals to find out who needs you.
- A CLI and a local socket API let agents drive Herdr themselves: split panes, start other agents, prompt each other.
- Local machines and SSH-connected machines appear side by side in one window, each reconnecting independently.
- The core is Apache-2.0 on GitHub. September 2026 brought a six-million-dollar seed round led by Bessemer Venture Partners, with Y Combinator participating.
- Current state: version 0.9.1, stable on macOS and Linux, Windows in beta. “Herdr Cloud” is announced but not yet available.
The Problem Behind the Tool
The tooling landscape for AI-assisted development has shifted fast over the past two years. Claude Code, Codex, Cursor, opencode, Grok — they all run in a terminal, they all work on a single task for minutes or hours, and they are all bound to the lifetime of their terminal process.
For one agent, tmux or screen solves this. Anyone who has worked on remote machines for a while does exactly that. It gets uncomfortable the moment several agents run at once — three projects, four agents, two of them waiting ten minutes on a question nobody has seen.
Classic multiplexers know nothing about their content. To tmux, a pane is a pane; whether an agent inside it is waiting for confirmation or running a test suite is irrelevant to the program. A human has to look. With four parallel agents, looking is the work.
What Herdr Does Differently
Herdr runs as a background server and manages the terminal sessions of every agent. Three things set it apart from a conventional multiplexer.
First: status detection. Herdr reads the output of each pane and infers the state of the agent inside it. The interface shows, per pane, whether work is happening, whether the agent is blocked and waiting for input, or whether it is done. That shifts the developer’s attention from “where do I stand everywhere?” to “who needs me right now?”
Second: agents as the primary user group. Herdr ships a CLI and a local socket API through which agents can operate the program themselves. An agent can open a new pane, start a second agent in it, and hand that agent a task. It can wait until another agent is genuinely blocked rather than blindly polling. This is the foundation for multi-agent setups where not every intermediate step routes through a human.
The boundary matters here: Herdr does not wrap the agents and does not replace them. It owns their terminals. The agents keep running unchanged, with the same commands and the same config files as before.
Third: machines, not just sessions. Local machine, office workstation, cloud server — Herdr connects them over SSH and presents them next to each other in a single window. Each connection reconnects independently of the others. If the server goes down, the local panes are untouched.
Under the Hood
Herdr is written in Rust and ships as a single binary — no Electron, no Node layer, no runtime package that needs installing alongside it. For a program meant to sit in the background permanently while parsing the output of several terminals, that is a sensible call: resource use stays low and startup is immediate.
The interaction model follows what tmux and zellij users already know. The default prefix is ctrl+b; ctrl+b q detaches the session. Anyone arriving without multiplexer experience can use the mouse instead: click panes, drag borders. The documentation explicitly addresses both groups.
On top of that sits a plugin system for local executable workflow plugins with manifest actions and event hooks. A marketplace is announced; publishing currently runs through GitHub.
Installation
Installation is a one-liner:
curl -fsSL https://herdr.dev/install.sh | shOr through the usual package managers:
brew install herdr
mise use -g herdrWindows has a PowerShell equivalent. Building from source needs nothing but a Rust toolchain:
git clone https://github.com/herdrdev/herdr
cd herdr
cargo build --releaseStart it with herdr. The same command reattaches to an existing session later.
Four Use Cases That Pay for Themselves
The overnight job. A migration task that takes two hours keeps running on the server while the machine at home is switched off. The next morning Herdr shows the state — finished, blocked, or crashed — without anything needing a restart.
Several projects in parallel. Three repositories, one agent per repository working on its own task. Instead of cycling through three terminal windows to check whether anything happened, a single overview answers the question.
Switching devices. Start on the office workstation in the morning, continue from the laptop in the afternoon — same session, same state, same scrollback history. No re-reading of context.
Agent starts agent. An orchestrator agent breaks a larger task into subtasks, opens a pane per subtask, starts a specialised agent in each, and collects the results. Without the socket API this would require writing your own process manager.
For teams already running AI agents in production — in maintenance work, say, or on recurring refactors — the fourth point is the interesting one. It moves the line of what can be automated without a human in the loop.
Funding, Licence, and the Obvious Question
Herdr was started by Can Celik. In early September 2026 the project announced a six-million-dollar seed round led by Bessemer Venture Partners, with Y Combinator, e2, and several angel investors also participating. The project had joined Y Combinator’s Fall 2026 batch shortly before.
With any open-source tool backed by venture capital, the same question comes up: what stays open? The project’s answer is stated plainly — the core remains free under Apache-2.0. The money goes to people, and to building a connected experience across machines plus a broader ecosystem of clients and plugins.
The commercial component is already visible in outline: Herdr Cloud is announced as an alternative for connecting machines without setting up SSH yourself. That is the familiar pattern — open core, paid infrastructure around it. Anyone using the tool purely locally is unaffected for now.
What to Know Before You Adopt It
The version number is 0.9.1. The project is young, and the API surface may still move. If you build the socket API into your own automation, expect to adjust it.
Windows support is marked beta and can collide with endpoint protection software. For Windows workstations under strict security policy that is a real obstacle — WSL or a Linux VM is the safer route there.
And Herdr solves a problem that only starts to hurt above a certain number of parallel agents. If you run one agent on one machine, tmux will do. The payoff begins with the third or fourth agent and grows with every additional machine.
Conclusion
Herdr answers a question that has come up more and more over the past months: where do the agents actually run when nobody is watching? The answer — a persistent background server that owns terminals, reads state, and can be driven by the agents themselves — is obvious enough that you wonder why it didn’t exist already.
Choosing not to replace existing agents but merely to take over their terminals is the right call. It keeps the barrier to entry low and makes the tool independent of whichever agent turns out to be most popular next year.
If you are working with more than two coding agents at once today, give it half an hour. Installation is one command, the licence is Apache-2.0, and the worst case is that you go back to tmux.
Sources: herdr.dev · Documentation · GitHub: herdrdev/herdr · Herdr blog on the seed round
