# Herdr turns the terminal into an agent control room > Source: > Published: 2026-07-08 04:43:22+00:00 Can Celik's [Herdr](https://herdr.dev/) is making a direct bet on where AI coding work actually happens: inside long-running terminal sessions, across several agents, with a human checking which one is blocked and which one is done. The July 2nd public push is simple on the surface. Herdr calls itself an "agent multiplexer" and gives developers one terminal for Claude Code, Codex, OpenCode, Kimi, GitHub Copilot CLI, Cursor Agent CLI and other command-line coding agents. Each agent runs in a real terminal pane. A Herdr server keeps the panes alive after the laptop closes, and the user can reattach locally, over SSH, or from a phone. The public repository identifies the maintainer as [Can Celik](https://github.com/ogulcancelik), whose GitHub profile links to [oddbit.ai](https://oddbit.ai) and pins other small agent-adjacent projects, including Pi extensions and a Unity bridge for AI-driven editor control. That matters because Herdr is less a hosted AI product than a maintainer-led attempt to define a missing layer around coding agents: the runtime that keeps their terminals alive, names their state, and gives agents a control surface of their own. The [GitHub repo](https://github.com/ogulcancelik/herdr) shows the project has moved quickly. As of July 8th, GitHub listed 13.6k stars, 790 forks, 70 releases, and v0.7.3 as the latest release on July 7th. The repository is mostly Rust, and the README says Herdr is dual-licensed under AGPL-3.0-or-later for open source use, with commercial licenses available for organizations that cannot comply with AGPL terms. ### The product is a terminal layer, not another agent app Herdr's pitch cuts against the current wave of agent workbenches that wrap coding agents in desktop apps, dashboards, or web views. Its [comparison page](https://herdr.dev/compare/) frames the product around persistence, SSH access, direct attachment to an agent terminal, semantic agent state, and an API. The homepage adds the blunt operational promise: stable Linux and macOS support, a Windows preview beta, no Electron, no account, and no telemetry. That architecture is the point. Herdr does not ask developers to move out of Ghostty, Kitty, iTerm, Alacritty, tmux muscle memory, or an SSH box where the code already lives. It keeps the interactive terminal as the unit of work, then adds a sidebar that rolls up agent status across panes, tabs, and workspaces. The [docs](https://herdr.dev/docs/) are written for two groups at once. New multiplexer users get mouse-first panes, draggable borders, right-click menus, and a quick-start path. Tmux and Zellij users get a familiar `ctrl+b` prefix, persistent panes, detach, and reattach. The onboarding prompt also tells users to let an already-running coding agent read Herdr's guide and explain setup back to them, a small but telling design choice: Herdr treats agents as users of the tool, not just processes inside it. ### State detection is the core bet Herdr's most important claim is that it can tell whether an agent is idle, working, blocked, or done. The [agents guide](https://herdr.dev/docs/agents/) says Herdr first detects the foreground process in each pane, then relies on either lifecycle hooks or screen manifests depending on the agent. Claude Code, Codex, GitHub Copilot CLI, Droid, Qoder CLI, and Cursor Agent CLI are handled with screen-manifest detection for state, while agents with fuller lifecycle hooks can report state more directly. That distinction is more than implementation detail. The operational pain Herdr targets is the moment a developer has six agent panes open and one is waiting for a permission approval while another finished a test fix ten minutes ago. Herdr's sidebar rolls status up to the pane, tab, and workspace, so a blocked agent marks the larger workspace as blocked and a done agent remains visible until viewed. The docs also show the limits. Blocked detection is deliberately strict for screen-manifest agents. If Herdr does not recognize an unusual prompt shape, it can fall back to idle until the manifest knows that screen pattern. The product's answer is remote manifest updates from herdr.dev, local overrides, and an `agent explain` command for debugging detection. That is a pragmatic design, but it also means Herdr's usefulness depends on staying current with the constantly changing UIs of the agent CLIs it watches. The "no telemetry" positioning should be read alongside that manifest system. Herdr says it has no account and no telemetry, while the agents guide says the binary can check herdr.dev for remote manifest updates unless a user disables that behavior. Those are different trust questions: Herdr can avoid collecting usage data and still fetch rule updates that affect how it interprets agent screens. ### The commercial opening is in the control surface Herdr is already more than a pane manager. Its [socket API docs](https://herdr.dev/docs/socket-api/) describe a local API for scripts and agents that can create and focus workspaces, split and resize panes, read output, send input, report custom agent state, subscribe to events, wait for output or state changes, and install integrations. CLI wrappers sit on top of the same surface. That is where Celik's project starts to look less like a nicer tmux and more like infrastructure for multi-agent development. If a coding agent can spawn panes, read logs, wait for another agent, or attach to a named agent terminal, Herdr becomes a local coordination layer. It does not own the model, the IDE, or the code review flow. It owns the live terminal runtime where those pieces collide. The homepage lists Herdr as popular with individual engineers from companies including JetBrains, Docker, Vercel, Sentry, Google, NVIDIA, AWS, ByteDance, Tencent, Alibaba, Salesforce, IBM, Atlassian, Whop, and Automattic. Herdr is careful to call those individual engineers rather than company endorsements. The public pages do not disclose revenue, customers, funding, or a company structure around the project. That leaves the business model visible only in outline. The AGPL license creates a path for broad open-source adoption while reserving paid commercial licensing for organizations that want different terms. Sponsorship is also present in the repo. For now, the measurable signal is developer attention: a fast-moving Rust repo with thousands of stars, frequent releases, and a product surface aimed at a workflow that has become common faster than the tooling around it. Herdr's test is whether agent-heavy developers want a terminal-native runtime badly enough to install and keep a new multiplexer in the loop. The answer will not come from the homepage. It will come from whether blocked, working, and done become states developers expect their terminal to understand.