Orcabot

A sandboxed, multiplayer AI coding platform. Run Claude Code, Codex, Gemini, or a plain shell in the browser with zero setup — each dashboard gets its own isolated VM with a built-in Chromium browser, a secrets broker that keeps API keys away from the LLM, and a network egress proxy ("Little Snitch for AI agents"). Dashboards are Figma-like multiplayer canvases where terminals, notes, browsers, and integration blocks (Gmail, GitHub, Drive, Calendar, Twitter, …) are wired together with edges.

Two ways to run it

• Cloud — Cloudflare Workers control plane + Fly.io sandbox VMs, fronted by the Next.js dashboard. This is the hosted product.
• Desktop — a native Tauri app (desktop/) that bundles the entire stack locally: control plane on workerd, a local SQLite "D1" shim, and a sandbox VM via Apple Virtualization.framework (macOS) / QEMU (Linux). No cloud account needed at runtime.

Architecture

Browser / Desktop webview / orcabot CLI
        │
Cloudflare control plane (controlplane/)  ── auth, dashboards (Durable Objects),
        │                                     sessions, secrets, integration gateway,
        │                                     PTY tokens, billing, personal access tokens
Fly.io / local VM  sandbox (sandbox/)     ── PTYs, agents, /workspace filesystem,
                                              secrets broker, egress proxy, MCP server

The frontend never talks to a sandbox directly — all traffic flows through the control plane.

The orcabot CLI

The desktop crate ships a second binary, orcabot, that runs the whole stack headlessly and drives it from the terminal — an interactive TUI plus scriptable subcommands. It can switch a live session between surfaces (cli ↔ desktop ↔ web) and package a dashboard + workspace for transfer (export/import locally, push/pull against a remote control plane). See desktop/CLAUDE.md.

Repository layout

Path	What	Guide
frontend/	Next.js dashboard UI (canvas, blog, docs, admin)	frontend/CLAUDE.md
controlplane/	Cloudflare Worker control plane	controlplane/CLAUDE.md
sandbox/	Go sandbox server (PTYs, agents, FS, brokers)	sandbox/CLAUDE.md
desktop/	Tauri desktop app + the orcabot CLI	desktop/CLAUDE.md
bridge/	Node WhatsApp/messaging bridge (Fly)	bridge/CLAUDE.md
e2e/	Playwright end-to-end tests	—

Project-wide conventions and the security model live in the top-level CLAUDE.md.

Quick start

Desktop (everything local):

cd desktop && BUILD_VM=force sh scripts/build-desktop-resources.sh   # one-time: build VM + workerd + frontend
cd app/src-tauri && cargo build --release
# GUI:
VZ_CONSOLE_DIRECT=1 ./target/release/orcabot-desktop
# or CLI (boots the stack headlessly, opens the TUI):
./target/release/orcabot

Cloud (dev):

cd controlplane && npm run dev          # control plane on :8787
cd frontend && npx wrangler dev         # frontend on :8788
# sandbox: see CLAUDE.md "Running Locally" for the docker run command

Tests

• Sandbox (Go): cd sandbox && go test ./...
• Control plane (vitest): cd controlplane && npm test
• E2E (Playwright, against a live instance): cd e2e && ORCABOT_URL=… npm test
• Desktop regression smoke (live stack): desktop/tests/regression-smoke.sh