Comparison
Shelbi vs Herd
A desktop agent IDE for parallel local coding work, with git worktrees, source control, and Claude Code and Codex support.
At a glance
| Dimension | Shelbi | Herd |
|---|---|---|
| Form factor | Native macOS app, TUI, and CLI | Desktop GUI for Windows, macOS, and Linux |
| Concurrency model | Named workspace pool with reused worktrees | Agents spawned into separate worktrees |
| Where work runs | Local machines and SSH-connected hosts | Local agents, according to its privacy policy |
| Work intake | Configurable Kanban; local files or GitHub Issues | Agent spawn flow and Workspace TODOs |
| Automation | LLM orchestrator; optional Zen Mode | Agent chaining |
| Review flow | Workflow handoff and dedicated review workspace | Source Control and inline merge conflicts |
| Runners | Claude Code, Codex, configurable CLI runners | Claude Code and Codex; Kimi shown on homepage |
| Source | Open source, MIT | Proprietary redistribution terms |
| Pricing | Free software; agent usage separate | No public plan price listed; bring your AI accounts |
Checked November 4, 2026 against Shelbi v0.13.3 and Herd's public materials. Herd's download page and changelog still list v0.2.1 (May 2, 2026). No newer public release was listed when checked. This comparison uses those documented features.
Where Shelbi excels
A native window across machines
Shelbi now has a native macOS desktop app alongside its TUI and CLI. The site shows the desktop client by default. Both clients share the same application models and command registry, including the Issues board and workspace terminals.
There are two ways to work remotely. Your project can dispatch agents to workspaces on SSH-connected machines. Or you can open an entire remote project through Connect to Remote Session…. The desktop opens its full native window over one shared SSH relay channel per host. Both machines need the exact same Shelbi version; a mismatch opens the remote TUI in a terminal view instead. On macOS, SSH password and passphrase prompts use a native dialog.
Remote agents keep running if your viewer disconnects, provided their host keeps running. Herd describes local execution in its privacy policy; its public materials do not document equivalent remote-project access.
A task board with configurable workflows
Shelbi's board follows your workflow definitions rather than five fixed columns. Statuses declare who owns the next step, which agent runs it, and which transition actions follow. The board can stay in local markdown or use GitHub Issues, where each card is a real issue.
The New task dialog in both clients includes Title, Workflow, Status, Priority, Description, and Let Orchestrator write this up. The switch asks the running orchestrator to expand your draft into a task with acceptance criteria.
Herd's documented intake includes a spawn palette and TODOS.md entries that
can become GitHub, GitLab, or Gitea issues. Shelbi is the closer fit when you
want the issue board itself to drive dispatch and review.
An orchestrator you can configure
Shelbi runs an orchestrator agent that reacts to project events and dispatches work to available workspaces. Onboarding lets you choose its harness and default model, then edit its instructions. Agent instructions and skills are editable too. Runner and model settings resolve by agent and task, so mixing Claude Code and Codex does not require a fixed runner per workspace.
Zen Mode is optional and off by default. When enabled, it lets the orchestrator promote and merge eligible work under your policy. Workflow ownership controls where it can act; configured local checks, CI, conflict checks, and danger paths inform the merge decision. Individual issues can opt out. This is more setup than Herd's documented chaining, but gives you control over the path from intake to merge.
Explicit review and a record of what shipped
A worker signals completion with its task id in
<worktree>/.shelbi/shelbi-ready. Shelbi routes the handoff through the task's
workflow. The default flow stages finished work in Review, with a dedicated
review workspace, View Diff, Approve, and Reject controls. Custom
workflows can define other handoff steps.
The project's Changelog records shipped changes and
links them to issues and PRs. You can read it in either client or through
shelbi changelog list.
A CLI alongside the app
Scripts and agents use the same commands: shelbi issue add,
shelbi issue move, shelbi issue prio, shelbi workspace list, and
shelbi events tail. The desktop is another way to drive that system, rather
than the only entry point.
Where Herd excels
A graphical client on more platforms
Herd offers Windows, macOS, Linux .deb and .rpm downloads, plus documented
WSL support. Shelbi's native app requires an Apple silicon Mac running macOS 15
or later; its TUI is available on macOS and Linux. For a graphical Windows or
Linux workflow, Herd has the broader platform coverage.
Direct agent spawning and source control
Herd's palette asks for an agent type and workspace. Its Source Control panel, inline conflict resolution, and AI-generated commit and PR messages suit someone who wants to launch agents and inspect their changes directly. That can be a simpler fit than configuring an orchestrator and a workflow.
GStack integration
Herd advertises preloading skills such as GStack, and its live demo shows a GStack review pipeline. If you already use that workflow, the integration may matter more than Shelbi's configurable agent roles.
Choose Shelbi if…
- You want a native Mac app or terminal client controlling local and remote work.
- You want a configurable issue board to drive dispatch, review, and optional automation.
- You want editable orchestrator instructions, agent roles, and model settings.
- You need a CLI for scripts and other agents, and want MIT-licensed source.
Choose Herd if…
- You need a graphical client on Windows or Linux.
- You prefer spawning agents directly over defining a board-driven workflow.
- Inline conflict resolution and generated commit messages are central to your review process.
- You want its advertised GStack integration.