Your private agent desktop.
StarloOS runs inside your existing operating system. It is not a bootable OS or a bundled general-purpose model.
01 / Install the workspace
Windows x64: use the setup executable. It is unsigned and native Windows execution remains unverified. Do not disable protection to force installation. First setup downloads a checksum-pinned Node.js runtime after confirmation.
macOS and Linux: extract the portable ZIP and run the included shell launcher. This is not a signed .dmg or AppImage. The runtime download needs internet on first setup.
bash Start-StarloOS.sh
Developers: with Node.js 22.16+ in the 22.x line installed, no package installation is needed:
npm startChoose a download →
02 / Start without a model
Local storage and private access are created automatically. Open the workbench, analyze a CSV, compare documents, train a small classifier, or use Research Desk. The first-run tutorial calculates an explicit sample dataset; it is not a simulated generative AI response.
Switch apps from the dock or Ctrl/⌘ K. Move, resize, minimize and maximize windows. Tasks continue while the local process runs, even if you close their app window.
03 / Connect a model for open-ended work
Open Model connection. Use your OpenAI/OpenRouter account or a separately installed compatible Ollama model. No model weights or paid usage are included. Supply keys only in your local application. Remembered cloud keys use a passphrase-encrypted vault; the passphrase is not stored.
Create a mission, choose an agent or crew workflow, set the budget, and authorize the provider request. Review tool requests and deliverables in Mission control. Credits limit execution; they are not a STARLO balance or prepaid service.
04 / Pause, review, and stop
Control center pauses new dispatch. Already-running provider requests may finish and incur costs. Task manager lists real local and AI jobs, with inspection and cancellation. A completed execution still needs your assessment. Stopping StarloOS requires outstanding jobs to be resolved.
05 / Keep a private snapshot
Export a snapshot from Control center after resolving all tasks. It includes research, agent settings, memories, trained models and completed local jobs. API keys, login codes and other users are omitted. The content is unencrypted and may be sensitive.
Restore only trusted files into a new folder, while the runtime is stopped:
node desktop/backup.mjs restore snapshot.json /new/workspace/folder
Start using STARLO_DATA_DIR pointing to that folder. Restore begins paused and requires model reconnection. Hashes detect file inconsistency, not truth or publisher identity.
Verification is scoped—not implied.
Local HTTP acceptance, persistence, calculations, permission checks and snapshot restoration have automated tests. Model decisions are fixtures in automated integration tests. Native Windows execution, macOS bootstrap, real provider calls and live Vercel deployment are not certified. This is an early-access agent application, not a finished or audited token economy.
Read the current verification report →
DEX Screener, Buy and CA remain NOT LIVE until verified launch information is published.