Orb

All your AI providers, machines, and agent harnesses in one place.
Connect your subscriptions, API keys, and local hardware to your machines, and orchestrate every agent harness from macOS and iOS.
_{Formerly sandboxed.sh — sandboxed.sh is now the backend to the Orb clients.}

Get started · Documentation · Desktop · iOS · Website · Discord

![Orb on macOS, following a Verity coding mission](screenshots/orb-desktop.webp) ## Choose where your agent works | Mode | Where it runs | How you use it | | --- | --- | --- | | **Your computer** | A local coding agent on your desktop | Choose **New Agent → This computer**, then a harness and model. Work with your local files and tools. | | **Your private cloud** | Your own servers and remote machines | Add machines to sandboxed.sh, then select one in Orb. Run agents in the configured host or isolated container workspace. | | **Cloud agents** | A coordinator or provider-managed assistant | Choose **Cloud agent**, then Hermes, ChatGPT, Grok Bot or Cursor Cloud. Continue durable or provider conversations from Orb. | Use **Claude Code**, **Codex**, **Antigravity**, **OpenCode** or **Grok** for local and remote work, with live model, effort and cyber-access controls where supported. Cloud agents have their own connections: **Hermes** binds to durable coordinator conversations and router chains; **ChatGPT** uses a signed-in browser profile and your available subscription modes; **Grok Bot** uses its connected account; **Cursor Cloud** uses its official API with a repository, Git reference and model. ![Creating an agent in Orb on macOS](screenshots/orb-setup.webp) ## Pick up the conversation anywhere Projects organize missions, per-project colors, skills and a synced local **Context** folder in Finder. Inside a conversation: - **Follow and steer live runs**: collapsible `Worked — …` tool/thought folds, nested subagents and callback reviews, inline prompt editing, and a Cursor-style follow-up queue that sends on turn completion or immediately. - **Ask on the side**: open the docked `/btw` side-agent panel (`⌘⇧J`) to ask questions with shared workspace context without interrupting the main turn. - **Keep long work moving**: schedule durable mission wake-ups, automatically resume interrupted runs after restarts, and transfer workspaces across machines (preserving symlinks while skipping build artifacts). - **Rich output**: Markdown, code, tables, LaTeX, quizzes, inline images and clickable file links stay readable inside the transcript. Orb for **iOS** connects to the same sandboxed.sh server: browse projects with synced colors, inspect mission transcripts and activity folds, send follow-ups, view or edit shared Markdown context, and manage backend, machines and providers.

Orb on iOS showing the production projects list with synced project colors Orb on iOS showing the Pareto mission conversation in Verity
_{iOS Simulator · Connected to production (`Projects` list and `Verity / Pareto` mission).}

## Bring your subscriptions Sign in to supported subscription accounts directly from **Providers** in Orb (macOS, web and iOS) or connect them through **CLIProxyAPI Plus**. You can authenticate via OAuth, renew tokens, enable or disable individual accounts, and inspect live plan/quota status; the proxy handles routing and rotation across eligible accounts while Orb lets you choose the harness and model. - [CLIProxyAPI Plus — maintained CCS fork](https://github.com/kaitranntt/CLIProxyAPIPlus) - [CLIProxyAPI — upstream](https://github.com/router-for-me/CLIProxyAPI) - [Credential ownership and proxy configuration](docs/CREDENTIAL_OWNERSHIP.md) This routes **model inference** for coding agents. Managed cloud agents use separate service adapters; connecting a proxy account does not sign in to the ChatGPT browser or create a Cursor Cloud API account. ## Get started 1. **Run sandboxed.sh.** Follow the [Docker guide](docs/install-docker.md) or [native Linux guide](docs/install-native.md). It keeps the project record, mission history and remote execution services. 2. **Open Orb.** Build the [desktop client](orb/README.md) or the [iOS app](ios_dashboard/README.md), then enter your server URL and sign in. 3. **Connect your execution environment.** Set up local harnesses, register [remote machines](docs/REMOTE_NODES.md), or connect a cloud-agent account. Configure inference providers for the models you want to use. 4. **Create a project and launch an agent.** Choose its execution mode, write your task, and keep its conversation and results together. ## Under the hood **Orb is the client; sandboxed.sh (the project's former name) is now the backend to the Orb clients.** Desktop is built with Tauri and SolidJS; iOS uses SwiftUI. The Rust backend owns projects, missions, event history and remote-node coordination. Cloud adapters observe provider work independently of the client window. The same control plane is available over MCP for coordinators such as [Hermes](https://github.com/Th0rgal/hermes-agent). Automation uses the same project and mission records as Orb. [Documentation index](docs/README.md) · [Execution architecture](AGENTS.md) · [Workspaces](docs/WORKSPACES.md) · [MCP and Hermes](docs/HERMES_ORCHESTRATION.md) · [Development and troubleshooting](DEBUGGING.md) --- **Orb** (`orb.thomas.md`) · Formerly `sandboxed.sh` (now the backend to the Orb clients).