Orb – Unified AI Agent Hub

What it is

Orb is a desktop (macOS) and iOS application that lets you run and manage AI‑powered coding agents from a single place. It talks to a self‑hosted backend called sandboxed.sh (now just the server side) which stores project data, mission history and coordinates remote execution.

Core ideas

Idea	How Orb implements it
One UI for many agents	Choose a harness (Claude Code, Codex, Antigravity, OpenCode, Grok, etc.) and a model and run it locally, on your own cloud, or via a provider‑managed cloud agent (ChatGPT, Hermes, Grok Bot, Cursor Cloud).
Projects & missions	A project groups related work. Inside a project you create missions (tasks) that keep a transcript, a synced Context folder, colors, and skill tags.
Live, steerable conversations	The transcript shows collapsible “Worked — …” folds, sub‑agents, inline prompt editing, and a follow‑up queue that can auto‑send the next turn.
Side‑panel for quick questions	Press ⌘⇧J to open a docked /btw panel that can ask a question using the same workspace context without breaking the main flow.
Durable, resumable runs	Missions can be scheduled to wake up later, automatically resume after a restart, and be transferred between machines while preserving symlinks and skipping build artefacts.
Rich output	Markdown, code blocks, tables, LaTeX, quizzes, inline images and clickable file links are rendered directly in the transcript.
Cross‑device sync	The iOS client connects to the same sandboxed.sh server, showing the same project list, transcripts and context files, and lets you edit or follow‑up from a phone or tablet.

How you use it (high‑level workflow)

1. Run the backend – either via Docker or a native Linux install (sandboxed.sh). It stores all project/mission data and exposes an API.
2. Open the client – build the macOS Tauri app (or the SwiftUI iOS app) and point it at your backend URL.
3. Add execution targets –
   • Local: run a coding agent on the same machine.
   • Remote: register a remote machine (SSH address book) that sandboxed.sh can dispatch jobs to.
   • Cloud: link a provider account (ChatGPT, Grok, etc.) through the Providers UI or via CLIProxyAPI Plus.
4. Create a project → launch a mission – write a task description, pick the mode (local, private‑cloud, cloud‑agent), select the harness/model, and start the conversation.
5. Interact – edit prompts, approve sub‑agent suggestions, ask side‑questions, or let the system run automatically.

Supported AI back‑ends & models

• Local / private‑cloud harnesses – Claude Code, Codex, Antigravity, OpenCode, Grok.
• Cloud agents – Hermes (durable coordinator), OpenAI ChatGPT (via signed‑in browser profile), Grok Bot, Cursor Cloud (official API with repo + git reference).
• Proxy routing – You can attach subscription accounts through CLIProxyAPI Plus, which handles OAuth, token renewal, quota display and request rotation.

Architecture snapshot

• Client: Tauri + SolidJS for macOS, SwiftUI for iOS – thin UI layer.
• Backend (sandboxed.sh): Rust service that owns project records, mission history, and coordinates remote‑node execution. It also exposes the MCP control plane used by orchestrators like Hermes.
• Remote nodes: Either local processes, Docker containers, or SSH‑accessible machines that run the actual model inference.
• Cloud adapters: Separate modules that talk to provider APIs; they are independent of the UI.

Who might benefit?

• Developers who want a single pane of glass for multiple coding‑assistant models.
• Teams that run AI agents on private infrastructure (on‑prem servers, self‑hosted GPUs) but also want to fall back to provider‑hosted agents.
• Power users who need persistent, resumable missions across devices and want rich, searchable transcripts.

Where to learn more

• Docs – docs/README.md (index), AGENTS.md (execution architecture), WORKSPACES.md (shared context), HERMES_ORCHESTRATION.md (MCP & Hermes), DEBUGGING.md.
• Desktop build guide – orb/README.md.
• iOS build guide – ios_dashboard/README.md.
• Community – Discord link in the header.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.