Shelbi vs Conductor

At a glance

Checked November 4, 2026. Conductor's announcement changes the decision: new users can no longer sign up or download the app. Existing Pro users keep Pro features free; existing free users cannot upgrade. Cloud becomes read-only November 13 and shuts down November 20, when cloud data deletion begins. Local Conductor receives new models and critical bug fixes until early 2027, when it starts winding down. Existing cloud users should export before November 20, using Settings → Account → Export data, or move a workspace and its chats with Transfer to local.

Where Shelbi excels

A reusable workspace pool

Shelbi declares named workspaces in the project YAML. Each handles one task at a time and keeps its worktree at <machine.work_dir>/.shelbi/wt/<workspace-name>, switching branches between assignments. You choose the pool size and which machine owns each slot. That keeps checkout paths and build caches available across tasks.

Conductor organizes independent tasks around separate workspaces and branches. It also supports several agents sharing the same workspace when they need the same files and branch. Shelbi's pool is useful when you want a bounded queue of reusable workers; Conductor's model fits work you create and review one branch at a time. See Workspaces.

Your machines, through a desktop or terminal client

Shelbi has a native macOS desktop app as well as the TUI. Both use the same application model and command registry. The desktop app is now the site's default view, with Orchestrator, Issues, Changelog, workspace terminals and review controls.

There are two ways to work remotely. Your hub can dispatch tasks to workspaces on other machines. Or you can open a whole remote project from Connect to Remote Session…. With matching Shelbi versions, the desktop opens the full native window over one SSH relay channel per host, without port forwarding. Different versions fall back to the remote TUI in a terminal view. Password-only SSH hosts work too, with a macOS sign-in dialog.

Remote agents keep running on their host when your viewing connection drops. The host must stay running. Conductor's managed cloud workspaces and live team collaboration are subject to the wind-down dates above. Shelbi's remote execution uses machines you control.

An orchestration loop you can edit

Shelbi's orchestrator can run on Claude Code or Codex. Its onboarding lets you choose the harness and default model, then edit its instructions before launch. Agent settings and task overrides select worker runners and models, so a workspace can run different providers across assignments.

The orchestrator's prompt lives at <config-root>/agents/orchestrator/instructions.md, with shared context in agents/_shared/preamble.md. The config root is normally ~/.shelbi/projects/<name>/, or <repo>/.shelbi/ in in-repo mode. Workflow YAML declares statuses, owners, agents and transition actions. You can inspect the implementation as well as change the policy. See Orchestrator and Workflows.

Conductor also exposes configurable repository prompts and actions. Shelbi adds access to the orchestration code and project-level workflow definitions.

A diff view you drive from the keyboard, in both clients

View Diff opens Shelbi's own diff view in the desktop app and the TUI, over a review branch or a workspace's live worktree. It has unified and split layouts, one commit at a time, syntax and word highlighting, search across files, folds, a blame gutter and viewed marks. Its keys follow Skim's and are rebindable.

Line comments are threaded. gY sends them to the workspace's agent, which can answer in the thread, and in a review it starts a Reject with them as the reason. In a workspace, gr and gR revert a hunk or a file, undoing committed work as an uncommitted change. On a task with a PR, the view shows its checks and reviews, and GitHub's threads at their lines, where you can reply, resolve and submit a review.

Conductor's viewer covers much of the same ground, with unified diffs, commit filtering, line comments and GitHub threads. What's different here is the terminal client, split view, the keyboard model and revert from the diff, plus another tool one setting away if you prefer it: diff_tool and review.diff_command still open an external diff tool.

A board, review step and shipping record

Shelbi can keep its issue board in local files or use GitHub Issues as the tracker. The New task dialog has Title, Workflow, Status, Priority and Description fields, plus Let Orchestrator write this up to expand a draft into a task with acceptance criteria. Both clients also have a project Changelog page for shipped work.

In the default flow, a worker writes .shelbi/shelbi-ready to hand off its branch. Finished work reaches Review, where you can load a dedicated review workspace and Approve or Reject it. The events log records transitions as text.

With Zen off, the default workflow waits for your approval. Zen Mode can promote backlog work and land finished branches under configured policy. Configured local checks, CI, conflict and sensitive-path checks form part of that process. Workflow ownership and per-issue settings control what can be automated. Keep Zen off when you want to initiate those decisions yourself.

Where Conductor excels

These strengths remain useful to existing users during its wind-down.

A guided path from review to merge

Conductor's diff viewer supports unified diffs, filtering by commit and comments on changed lines that go back to the agent. Around it, the Checks tab collects git status, PR details, CI, deployments, GitHub review threads and todos, and can send unresolved GitHub comments to the agent. Conductor also suggests the next step as a workspace nears merge, such as creating a PR, fixing checks or merging, and can discourage a merge while required work is open.

Shelbi's diff view shows a PR's checks, reviews and threads, but not deployments or todos, and gY sends your own comments rather than GitHub's threads. It doesn't suggest what to do next: opening and merging a PR are workflow actions that run on a transition, so you set them up once rather than choosing them from the review.

Session-level agent and model controls

Conductor supports Claude Code, Codex, Cursor and OpenCode. It bundles Claude Code, Codex and OpenCode; Cursor runs through its API. Each chat selects a harness and model, and several chats can share a workspace. Models depend on the account and provider: Claude models through Claude Code, GPT models through Codex, Composer 2.5 through Cursor, and provider-supplied models through OpenCode. See Harnesses and Agent modes.

Recent releases added Haiku 5.5 and GPT-6.1 Terra. Plan Mode, Fast Mode, reasoning controls and checkpoints vary by harness. Shelbi supports runner and model configuration too, but Conductor puts these choices directly in the chat interface.

GitHub and Linear context at workspace creation

Conductor can start a workspace from a branch, PR, GitHub issue or Linear issue, then guide it through review, PR creation, checks and merge. See Your first workspace. For an existing user whose work starts in Linear and finishes in GitHub, that connection remains useful. Shelbi now has a GitHub Issues backend, so GitHub integration alone is no longer a reason to dismiss it.

Workspace actions and preview URLs

Conductor's Run button launches saved actions inside a workspace. CONDUCTOR_PORT supplies workspace-specific ports, and configured preview URLs appear in the terminal panel's Open menu. Spotlight testing can sync a workspace back to the repository root for projects that must run there. See Testing and Settings reference.

An iPhone companion

Conductor's October releases added iOS access and remote control of local Mac workspaces. You can read chats, send prompts, select models and stop agents from your phone. Remote control is experimental and requires Conductor to remain open on an awake, online Mac. It connects through Conductor's relay; this is different from Shelbi's SSH access to your own hosts. See Remote Control. The announcement gives no separate long-term commitment for that relay.

Choose Shelbi if…

• You want an open-source orchestrator you can install and run on your own machines.
• You want a native Mac app or terminal client for a queue of persistent workspaces, including workers over SSH.
• You want to open a remote project's full desktop interface while its agents keep running on the remote host.
• You want editable workflow policy, a dedicated review step and optional Zen automation, with either local files or GitHub Issues as the tracker.
• You want to review in a keyboard-driven diff view, in a terminal or the Mac app, with comments that reach the agent and revert in place.

Choose Conductor if…

New users cannot choose it today: sign-ups and app downloads are closed. For existing users, continuing temporarily may make sense if:

• Its Checks tab, model picker or Linear-to-workspace flow is central to your current work.
• You want to finish work in the local app while it still receives model updates and critical fixes, through early 2027.
• You are transferring cloud workspaces to your Mac and exporting data before the November 20 shutdown. Plan around the published wind-down rather than adopting Cloud for new long-running work.