Short answer: Herdr is worth a pilot if you already run CLI coding agents such as Claude Code, Codex, or OpenCode and want their sessions to survive a closed laptop, a dropped network, or a machine restart. It is not a hands-off production system yet: this is an observed source and documentation review, the newest build is a prerelease, and open issues touch common workflows. Evaluate it in a non-critical workspace before standardizing on it.

Key facts at a glance

What changed in the 2026-08-18 preview

Per the release notes published by the maintainers, this preview is built from commit 9fac51722653 on master, with stable v0.8.0 as its base. The main change is that Windows is now generally available. The build also adds plugin marketplace discovery shelves for trending and new arrivals, fixes macOS IME commits in report-all mode, isolates hidden pane render cadence, retries Windows installer activation after transient locks, reaps ctrl-click URL openers, and publishes preview documentation.

Because this is a prerelease, read the changes as direction rather than a final contract. The previous preview from 2026-08-17 and the v0.8.0 stable release carry much larger change sets, including the relicensing to Apache-2.0 and the bundled Windows ConPTY runtime that arrived with v0.8.0.

What Herdr actually does

The maintainers describe Herdr as “the runtime your coding agents live on”: a background server written in Rust where terminals live inside it. Agents keep working when you close the lid, drop the network, or restart the machine, and sessions come back; you can reattach from any terminal or over SSH. Each pane is marked working, blocked, or idle, so you do not have to hunt for the stuck agent.

Per the README, Herdr does not wrap or replace Claude Code, Codex, Cursor, OpenCode, Grok, or other tools; it owns their terminals. Agents can drive Herdr through the CLI and socket API, spawning panes, prompting each other, and waiting until another agent is blocked. Plugins extend panes and workflows through a marketplace, and it ships as one Rust binary with no Electron. The README lists a curl installer, Homebrew, mise, and a PowerShell installer for Windows.

Who should consider Herdr

Herdr makes sense when you already depend on one or more CLI coding agents. The persistence and status model solves a real problem for anyone running agents on shared or remote machines, where a dropped SSH connection can kill a long job. It also suits teams building agent-to-agent workflows that want a socket API and plugin surface instead of screen-scraping terminals.

It is a weaker fit if your team does not use CLI-based coding agents, if your agents run in fully managed web interfaces, or if you want a hosted platform rather than software you install and operate yourself.

What to check before production

• Pin a version. The newest build is a prerelease. Start from stable v0.8.0 for production, and verify the stable release checksums the maintainers now publish, as noted in the v0.8.0 release notes.
• Test agent detection in your exact environment. Detection is heuristic. Open issue #803 reports that a nix-wrapped Claude Code binary (.claude-unwrapped) is never detected in devenv or nix shells, leaving agents invisible in the panel; the reporter worked around it with an alias.
• Check clipboard and terminal behavior. Open issue #2399 reports copy failing silently when the outer terminal ignores OSC 52, such as VS Code’s integrated terminal over SSH or WSL; the reporter found herdr --remote avoided the problem.
• Treat Windows as new. Windows GA arrived in this preview, and the same release notes list a batch of Windows fixes; pilot on Windows before relying on it.
• Decide who owns the daemon. Persistent sessions mean supervised background processes on the host, including upgrades, restarts, and cleanup.
• Review remote attach. Reattach over SSH is a feature and a security surface; confirm authentication and host access before exposing it to a team.

Maintenance signals

Herdr is actively maintained but very young: the repository was created on 2026-03-27, and the last push was the day before this review. At review time, GitHub showed 30,367 stars, 2,163 forks, and 197 open issues. The maintainers’ release notes describe restricting unapproved pull requests to bug fixes and gating review by scope, which indicates a deliberate contribution process rather than an unmoderated one.

The license change in v0.8.0 from AGPL-3.0-or-later to Apache-2.0 matters for commercial use: redistribution and integration are simpler under Apache-2.0.

Limitations and honest caveats

This review is a source and documentation review; no local installation or runtime verification was performed, so claims about persistence, status detection, and performance are the maintainers’ descriptions, not independently measured results. The project is only a few months old, so API stability, plugin marketplace governance, and Windows maturity are still unproven. Agent support depends on detection heuristics that track each agent’s UI, and the changelogs contain many fixes in that area, so expect occasional breakage when agents ship new versions.

Evaluation questions before you adopt

• Which agents does your team run, and are those exact versions in the supported list?
• Do your developers use nix, devenv, WSL, or VS Code integrated terminals? Test those combinations first.
• Can you stay on the stable channel while the preview channel matures?
• Who handles upgrades, monitoring, and incident diagnosis on the host machines?
• Do you need the socket API and plugins now, or would plain session persistence be enough?

Verdict

Herdr addresses a real operational gap: coding agents are long-running interactive processes, and a multiplexer built around their behavior is more useful than a generic tmux setup. The Windows GA preview and plugin marketplace are meaningful progress. But the recommendation is conditional: pilot it on staging, pin a version, verify checksums, and test your exact terminal, OS, and agent combination before letting it own production sessions. If those checks pass, Herdr is a credible option in this space; if they fail, you avoided a messy rollout.

Sources

• herdrdev/herdr repository
• Preview build 2026-08-18 release notes
• v0.8.0 stable release notes
• Issue #803: nix-wrapped Claude Code is not detected
• Issue #2399: clipboard copy fails when the outer terminal ignores OSC 52
• Issue #1803: tab name dialog navigation
• Herdr documentation
• Herdr quick start

What I can do next

If this evaluation matches your situation, the practical next step is a scoped pilot. At XP812 I work on AI agent deployment and integration: setting up Herdr on the Linux or macOS machines your agents actually run on, wiring it to the agents and plugins you already use, and then taking over the production side with pinned upgrades, monitoring of long-running sessions, and diagnosis when agent detection or terminal behavior misbehaves. That is deployment and integration now, and maintenance afterward.