Herdr

Herdr is one Rust binary that acts as a session server plus terminal clients. The server owns the terminals (panes), grouped into tabs and workspaces, and keeps them alive when clients detach. It does not wrap or replace the coding agents: the agent runs as an ordinary process in a pane, and Herdr recognizes it from the foreground process, screen-detection manifests, or an installed integration that reports state.

Coordination is done by whoever drives the CLI or socket API. The docs describe three primitives: layout (workspaces, tabs, pane splits), pane (raw terminal control) and agent (a recognized coding agent addressed by name). An agent with the skill file installed can create panes, start other agents, prompt them and wait until they are idle or blocked. Herdr itself does not assign tasks or merge results.

State is the server's live process tree plus a saved layout. After a server or machine restart the layout returns but the original processes do not; agents whose integration reports a resume command can pick up the same session. Herdr does not call any model, so it does not proxy model traffic.

Agents run in real terminals, so a person can attach, read and type at any time. Herdr marks a pane blocked when the visible screen shows an approval, question or permission prompt, and notifies the user. Scripts can send text or keys to an agent. The docs say a wrong status guess only affects the displayed status and waits, and should not make Herdr send input or take destructive action. There is no built-in diff review or approval step; that stays in the agent and in the user's own git tooling.

curl -fsSL https://herdr.dev/install.sh | sh

curl -fsSL https://herdr.dev/install.sh | sh   # or: brew install herdr
herdr                                           # start or attach to the background session

# Scripting: create a workspace, split a pane, start Codex in it, prompt it
created=$(herdr workspace create --cwd ~/project --label api --no-focus)
pane_id=$(printf '%s\n' "$created" | jq -r '.result.root_pane.pane_id')
split=$(herdr pane split "$pane_id" --direction right --no-focus)
review_pane=$(printf '%s\n' "$split" | jq -r '.result.pane.pane_id')
herdr agent start reviewer --kind codex --pane "$review_pane"
herdr agent prompt reviewer "Review the current diff" --wait --timeout 120000

herdr worktree create --cwd ~/project --branch review   # git worktree as a workspace
# detach with ctrl+b q; run `herdr` again to reattach; `herdr server stop` ends the session

• Detaching keeps agents running; herdr server stop ends the session and its panes.
• After a server or machine restart the layout is restored but the original processes are not; agent session resume needs the agent's integration (herdr integration install <name>).
• agent start needs an existing pane sitting at a shell prompt.
• Blocked detection is strict: unusual new prompts may show as idle, and Codex can stay unknown, so waits for idle may time out (see the Agents page).
• Native Windows is supported but some capabilities differ or are unsupported (for example herdr terminal attach); the installer is a PowerShell command.

Nothing in the README or docs describes MCP support, so the record marks it unknown. Herdr's own control surface is a CLI and a socket API.

The Agents page lists Claude Code, Codex, OpenCode, Hermes Agent and many more as supported, and any other program runs as a plain terminal. OpenClaw is not on that list.

Yes. It is Apache-2.0 licensed. The README lists an enterprise and partnership contact address but no paid plan.

It can create and open worktrees as workspaces with herdr worktree create, but using them is up to you or your scripts; there is no automatic one-worktree-per-agent rule in the docs.

Herdr is a persistent terminal multiplexer with agent state tracking and a scripting API, and it runs any program. Claude Squad is a TUI that starts each agent session in its own tmux session and worktree and adds diff and commit shortcuts.