Herdr: manage coding agents in persistent terminals
Set up Herdr TUI for parallel coding agents. Understand SSH sessions, restart recovery, the separate native app and the model costs behind each pane.
Last updated: 2026-11-14
Herdr TUI runs coding agents in terminal panes and shows which ones are working, blocked or idle. Use it when you already work with tools such as Claude Code, Codex or OpenCode and want to follow several sessions from one terminal. Each agent keeps its own tools, model access and permissions.
The TUI is Apache-2.0 open-source software, free to use. Model calls still belong to the agent and provider you configure. Adding panes can increase concurrent usage and spend even when the terminal manager itself costs nothing.
Choose Herdr TUI or the native app
Herdr now also offers a native app built on herdr-core. This guide covers the terminal project at herdrdev/herdr. The app and TUI have separate sessions; a session created in one does not appear in the other. Choose the interface before following an installation guide or looking for an existing conversation.
The TUI provides persistent terminals, a combined agent list and local or saved SSH machines. Agent-aware status depends on an integration: a program can run in a terminal without Herdr recognizing when it is waiting for approval. Check the supported-agent list for the tools you actually use.
Install and start the terminal runtime
The official installation guide provides stable binaries for Linux, macOS and Windows. On macOS, Homebrew is one installation option:
brew install herdr
cd /path/to/your/project
herdr
For other platforms, use the official TUI installation instructions. Install and authenticate your chosen coding agent separately. Herdr does not supply an agent subscription or API credentials.
In a Herdr pane, run your existing agent command, such as claude or codex. Give it a small task with a clear completion condition, then open another pane for independent work. The quick start explains pane controls and navigation. Review each agent’s approval prompt in its own terminal.
For parallel changes to the same repository, give each agent a separate branch and working directory, for example with Git worktrees. Multiple panes opened in the same checkout share files and can overwrite each other’s edits. Herdr’s status view helps you notice waiting agents; repository isolation and conflict resolution remain part of your workflow.
What survives a disconnect or restart?
| Event | What to expect |
|---|---|
| Detach from the client | The background server keeps the original terminal processes running. |
| Lose a client SSH connection | Processes owned by the remote server can continue; reconnect to that server. |
| Restart the Herdr server or machine | Saved layout can be restored, but the original processes have ended. |
| Resume an agent after restart | Requires that agent’s supported conversation-resume integration. |
What survives each disconnect event and the parallel-agent budget example. Original diagram of documented behavior, not a recorded session.
Use ctrl+b q to detach, then run herdr to reattach. A restored pane layout
does not establish that an interrupted build, shell command or agent task
completed. Check its output and repository state after reconnecting. See the
session-state guide for the resume
behavior of supported agents.
Work with a remote machine
Configure an SSH target such as workbox and verify that ssh workbox can connect. Save it using the connecting-machines guide, then attach to that target:
herdr --remote workbox
Here, workbox is an example configured SSH alias or saved profile target. The remote Herdr server owns its panes and running processes. The machine needs the agent tools, repository access and provider credentials used by those tasks; installing the client on your laptop does not configure the remote environment. Follow the SSH guide for server setup and reconnection.
Keep the host budget separate from model usage. A rented machine can remain billable while its agents are idle or waiting for approval. Detaching a terminal client does not stop that rental.
Budget parallel agents by completed work
For API-backed agents, add each session’s fresh input, cached input, output and separately billed tool usage. For subscription-backed tools, inspect their current allowances and overage behavior. A subscription price does not translate directly into a per-token API rate.
As an illustrative budget, three agents running for 30 minutes on one $0.50/hour machine use $0.25 of machine time. If each session incurs $0.80 in model charges, total spend is $2.65 before other services. If only two tasks are accepted, that is about $1.33 per accepted task. These assumed figures explain the calculation; they are not Herdr or model-provider tariffs.
Parallel work can shorten elapsed time while increasing duplicated reading, competing edits and review. Start with independent tasks, keep the provider’s usage records, and compare the total bill with work you retain. Use the AI running-cost calculator when a separately operated model server is part of that budget.
Herdr versus DeepSeek Harness
Choose Herdr when your main problem is organizing existing agent terminals, spotting blocked sessions or reconnecting to remote work. Choose DeepSeek Harness when you want to configure the agent environment itself: model providers, workspaces, tools, sessions and plugins. Their costs follow the model services and infrastructure you select.
For native desktop actions, agent-desktop supplies a different capability: observing and operating application controls. Browse the Agent guides for other workflows.
Sources
Checked 2026-11-07: Herdr website and app/TUI FAQ, TUI repository and Apache-2.0 license, installation, quick start, agent integrations, session recovery, remote connections and saved machines.