Shelbi vs Herdr

At a glance

Both tools keep agents running on your machines and connect over SSH. Shelbi organizes work around issues and workflows. Herdr organizes the workspace around live agents and an interface you can reshape.

Herdr's app FAQ explains the split: the app and TUI do not share running agents. This comparison covers both, with TUI-specific features called out explicitly. Checked November 5, 2026.

Where Shelbi excels

An issue queue connected to the work

Shelbi's Issues board tracks work before an agent starts and after it finishes. The default task workflow runs through Backlog, Todo, In Progress, Review and Done, with Canceled also available. Workflows can add their own statuses, agents and transition actions.

The New task dialog in both clients sets Title, Workflow, Status, Priority and Description. Let Orchestrator write this up asks the orchestrator to expand a draft into a task with acceptance criteria. The board can use local markdown or GitHub Issues, so a card can also be an issue your team reads and comments on in GitHub.

Herdr can show a board too. Its guide includes agent cards grouped by state and opening a worktree for an issue. Shelbi's advantage is the supplied issue-to-delivery workflow: queue, dispatch, handoff and merge actions are connected from the start.

A project orchestrator with persistent capacity

Shelbi's orchestrator dispatches ready issues to named workspaces, each with a persistent git worktree and session on its assigned machine. The issue's prefers_machine field can constrain dispatch to a particular host. The pool makes available capacity explicit across local and SSH machines.

The orchestrator is no longer limited to Claude Code. Onboarding lets you choose its harness and default model, then edit its instructions before launch. Claude Code and Codex have dedicated adapters; other CLI runners are configurable, with fewer integration guarantees.

Herdr also supports agent-led coordination. Its core lets an agent launch helpers and hear back from them. The choice is between Shelbi's supplied project scheduler and composing coordination around Herdr's runtime primitives.

A review workflow with an explicit human handoff

With Zen Mode off, finished work waits for your decision. The default workflow pushes the branch and opens a PR on the move into Review; moving to Done runs the merge action. Review workspaces let you load the task, inspect the diff and run the project's review recipe before accepting or rejecting it.

Herdr's reviewer-helper example can get another agent's findings back to the parent. Shelbi provides the issue status and human review path around that kind of work as part of its workflow model.

Optional autonomy under a project policy

shelbi zen on, shelbi zen pause and shelbi zen off control Zen Mode. The project's editable zenmode.md instructs the orchestrator when to promote backlog and finalize work. The shipped policy checks local results, conflicts, diff size and danger paths, then requires a review against the issue and checks CI before moving passing work to Done. The workflow transition owns the merge.

The shipped policy leaves tasks routed to review workspaces for a human. Zen is an agent policy supported by CLI primitives, so its prose rules depend on the orchestrator following them. Herdr supplies coordination and customization tools; Shelbi ships this project-level policy.

Desktop and terminal views of the same project

Shelbi's native macOS app shipped in v0.11.0 and shares its application models, command registry and sessions with the TUI. Remote workspaces open from the sidebar with shell, diff and editor sessions. Since v0.13.2, a remote project can open in the full native window over one SSH relay per host. Both machines need matching Shelbi versions; otherwise Shelbi opens the remote TUI in a terminal view with a notice. v0.13.3 extends the native connection to password-only hosts.

The project's Changelog page, added in v0.12.0, gives both clients a searchable record of what shipped. See the release history for these changes. Herdr's app and TUI are separate runtimes, so switching between those two interfaces does not expose the same running agents.

Where Herdr excels

A native app across more operating systems

Herdr offers desktop builds for Apple silicon macOS, Linux on x86_64 and arm64, and Windows on x86_64. Shelbi's native app requires an Apple silicon Mac with macOS 15 or later; its terminal client supports macOS and Linux. If you want a graphical workspace on Linux or Windows, Herdr has the broader offering.

An interface your agent can rebuild

Herdr's app documentation describes layouts and JavaScript mods for panels, menus and dialogs. Personal files reload when saved. You can ask an agent to build a board, add a timer or change the workspace, and share layouts and mods.

The original TUI has a separate plugin marketplace and executable plugin system. Those are distinct extension systems. Herdr gives you more scope to invent a work surface; Shelbi supplies an opinionated board and workflow with settings you can edit.

Direct coordination primitives for agents

herdr-core provides hc run, hc send, hc read and hc wait for starting, steering and inspecting helpers. Its API is available to other clients, with hc schema describing requests and responses. This is a useful foundation if you want to build your own coordination tool.

Core status comes from programs reporting through OSC 7501. Programs without that protocol still run, but have no reported agent status. The original TUI additionally detects agents such as Claude Code, Codex, OpenCode, Pi and GitHub Copilot CLI from screen rules or integrations. That TUI support list does not imply the same status support in the new app. Models come from the agent you run.

A multiplexer without adopting a project workflow

The Herdr TUI keeps keyboard prefix controls alongside clicking, dragging and splitting panes. It fits a workflow where you want to arrange terminals and keep running agents reachable without first defining a dispatch pool or task lifecycle.

Persistence has limits worth understanding. Closing a viewer leaves programs running. After a reboot or core crash, the new app restores kept panes with fresh shells, not running agents. The original TUI can resume supported agent sessions from saved state; the old processes do not survive a restart.

Choose Shelbi if…

• You want issues, priorities, dispatch, review and merges in one workflow.
• You want a declared pool of workspaces across your machines, with per-task host routing.
• You want the macOS app and terminal client to access the same sessions, including a native window for a remote project.
• You want a supplied Zen policy you can tune, while keeping human review for tasks routed to review workspaces.

Choose Herdr if…

• You want a native agent workspace on Linux or Windows as well as macOS.
• You want agents to reshape the interface with layouts and mods, or you want the original TUI's plugin system.
• You want a terminal runtime and API for composing your own agent coordination.
• You mainly need live terminals and agent visibility, and prefer to arrange the issue and delivery workflow yourself.