Comparison
Shelbi vs Cursor Background Agents
Cursor's Background Agents are now called Cloud Agents. They run asynchronous coding tasks in managed environments or execute tools on self-hosted machines, with Cursor's cloud running the agent loop.
At a glance
Checked November 4, 2026. Shelbi capabilities below are available in v0.13.3.
| Dimension | Shelbi | Cursor Background Agents (now Cloud Agents) |
|---|---|---|
| Form factor | Native macOS app and terminal UI over the same sessions | Desktop, web, and native iOS app; Android via web/PWA |
| Where work runs | Local or SSH machines you operate | Managed Ubuntu VMs, or self-hosted tool execution with a cloud agent loop |
| Concurrency model | Configurable pool of named workspaces | Parallel agent runs on managed VMs or self-hosted workers |
| Environment | Persistent git worktrees across tasks | Prepared Builds for hosted runs; existing checkouts for self-hosted runs |
| Orchestration | Your agent CLI, instructions, and workflow files | Cursor service; APIs and a separate Agent SDK for customization |
| Review flow | Workflow handoff, review workspaces, optional Zen automation | Pull requests, artifacts, and remote desktop testing |
| Model selection | Claude Code, Codex, or a custom CLI runner | Curated Cloud Agent models; context size selectable where supported |
| Source | Open source, MIT | Proprietary commercial service |
| Pricing | Free software; agent and machine costs are yours | New individual plans include Cloud Agents from Premium, $30/month before tax; usage terms vary by model and plan |
Where Shelbi excels
Persistent workspaces on machines you operate
Each Shelbi workspace keeps a worktree at
<machine.work_dir>/.shelbi/wt/<workspace-name> and switches branches
between tasks. Existing dependencies and build output can stay on disk.
The workspace persists; an agent's conversation need not carry over to
the next assignment.
Cursor also avoids repeated setup. Its Builds prepare repositories and dependencies before hosted agents start. Its Self-Hosted Machines can use existing checkouts. Shelbi's advantage is direct control of the workspace pool and orchestration through your own configuration.
Native desktop access to local and remote projects
Shelbi now has a native macOS app alongside the TUI. Both use the same application models and command registry. The desktop requires an Apple silicon Mac with macOS 15 or later.
A remote project opens in the full native window over one shared SSH relay per host, without port forwarding. Local and remote Shelbi versions must match; otherwise the desktop opens the remote TUI in a terminal view with a version notice. Password-only hosts are supported. Remote workspaces also expose the agent, Shell, View Diff, and Edit in your chosen editor from the sidebar.
An orchestration loop you control
Shelbi launches agent CLIs on your machines. You choose the runner and model, edit agent instructions, and define workflow statuses and transitions. Orchestrator onboarding walks through the harness, default model, and instructions before launch.
Cursor's self-hosted option keeps the checkout and tool execution on your hardware, but Cursor still runs planning and the agent loop. File contents and tool results needed by the agent leave the machine. Shelbi removes that required Cursor service from the loop; the CLI and model provider you choose still have their own data handling.
Review policy you can configure
Shelbi's review workspaces can load a finished branch and run a workflow's configured server for human inspection. The Review view provides diff and editor access plus approve and reject actions.
Zen Mode changes who handles eligible work.
The shipped policy lets the orchestrator review and finalize passing
handoffs when no review workspace is configured. It preserves the human
gate for review workspaces. Policy lives in the project's zenmode.md;
human sign-off is not an unconditional property of every Shelbi workflow.
Cursor also supports review before merging. Its PR artifacts and remote desktop are useful ways to inspect a change. The choice is about where you configure and conduct that review.
A shared board and a CLI for automation
Shelbi supports a local file board or GitHub Issues as the issue tracker. Both clients offer a New task dialog with Title, Workflow, Status, Priority, Description, and “Let Orchestrator write this up”. A per-project Changelog records shipped work.
Scripts can use shelbi issue add, shelbi issue move,
shelbi workspace list, shelbi events tail, and shelbi changelog.
Shelbi itself is free under the MIT license. You pay for your agent
services and compute; that does not make task costs fixed.
Where Cursor Background Agents excels
Managed work that continues with your laptop off
With Cursor-hosted agents, Cursor operates the VMs and their lifecycle. Your laptop can be offline. Shelbi can run work on an always-on remote host, but you operate that host. Cursor's self-hosted workers also need their own machines online.
Web, mobile, and collaboration integrations
Cursor offers desktop and web access, an iPhone and iPad app, and Slack, Linear, GitHub, and Bitbucket triggers. Android users can use the web app/PWA. These are useful entry points for teams delegating away from a development machine.
Its October 20 Agent SDK release also supports agents defined in code, evals, webhooks, and schedules, deployable to Cursor or your own infrastructure. Shelbi's CLI is not a unique automation surface.
Visual verification and larger environments
Cursor agents can operate a desktop and browser and produce screenshots, videos, and logs. You can take control of the remote desktop to test the software. Embedding artifacts in GitHub PR descriptions is opt-in. Shelbi's served review branches are useful, but it does not provide the same built-in artifact capture workflow.
The October 16 Cloud Agent update added shareable interactive canvases, a beta Environments API, and support for environments with up to 300 repositories.
A managed product with explicit usage controls
For new individual subscriptions, Premium is $30/month before tax and includes Cloud Agents. Free and Plus do not. Teams and Enterprise also include them. Legacy paid plans continue under their existing terms, so $20 Pro remains relevant to existing subscribers rather than new-plan pricing.
On current plans, included usage covers Cursor models; third-party models are billed separately. You can set a spending cap. Check Models & Pricing for your plan. Cloud Agents use a curated subset of models, with context size selectable where supported. Cursor documents a Max Mode upcharge for legacy request-based plans, rather than mandatory Max Mode for every current run.
Choose Shelbi if…
- You want a native Mac app or TUI for a persistent workspace pool on local and SSH machines.
- You want to choose agent CLIs and edit orchestration instructions and workflows yourself.
- You want a file or GitHub Issues board, configured review workspaces, and a project Changelog in one client.
- You prefer free orchestration software and managing your own compute and agent costs.
Choose Cursor Background Agents if…
- You want managed execution that keeps running while your laptop is off.
- Web and mobile access, chat integrations, and hosted automation matter to your team.
- Built-in browser verification, artifacts, canvases, or large multi-repo environments fit your work.
- You want Cursor's service to coordinate agents, including when tools execute on your own machines, and its plan and usage costs suit you.