Bring your station online.
One website. Your model account. Your private workspace.
Uploading the website publishes its interactive station, labelled guided sample, and generic downloads. Real tasks require the private configuration below. The visual preview is not silently connected to a model.
1. Deploy to Vercel
Upload the supplied Vercel ZIP to Vercel Drop. Keep the included build settings. The project uses Node.js 22.x, build command npm run build, and output directory public. The application route is /app.
Each drop creates a new project. Test its generated address before moving starloos.com from your existing project. Nothing in this package buys or changes a domain.
2. Connect private storage and owner access
In the project’s Storage/Marketplace area, connect Upstash Redis to Production. Check that both database settings appear under Environment Variables. Add a unique owner password.
| Environment variable | Value |
|---|---|
ADMIN_PASSWORD | Your own unique secret of at least 24 characters. |
UPSTASH_REDIS_REST_URL | The database REST URL. |
UPSTASH_REDIS_REST_TOKEN | The private database REST token. |
Never prefix these values with NEXT_PUBLIC_ or place them in frontend source. Keep them out of chat.
3. Connect a funded model account
The simplest setup uses OPENAI_API_KEY. The default agent model is gpt-4.1-mini. Change AGENT_MODEL to a model that your provider account supports with function calling. Provider charges come from your account, not from the allocation of StarloOS credits.
OPENAI_API_KEY=your-private-key # Optional override: AGENT_MODEL=gpt-4.1-mini
Alternatively, use OpenRouter with a tool-capable model:
AGENT_PROVIDER=openrouter OPENROUTER_API_KEY=your-private-key AGENT_MODEL=openai/gpt-4.1-mini
Do not select a model based on its name alone: verify its tool support, pricing, response format, and account access. Model-authentication testing in Connections does not execute a paid inference request.
4. Redeploy and run one small task
Redeploy after setting the environment variables. Open StarloOS → Connect runtime, and enter your ADMIN_PASSWORD. The owner workspace receives 1,000 operator-funded pilot credits once.
Select Prism and enter: “Calculate (48 + 12) / 3 using the calculator. Explain the result and create calculation.md.” Start with a 100-credit maximum. Inspect the file-write approval, approve it, and download the resulting file.
The operator pays actual model, search, storage, and hosting bills. Internal credits only limit service usage. They are not money, tokens, or a guarantee that a task will succeed.
Optional: web discovery
Set BRAVE_SEARCH_API_KEY for the search tool. Agents can read supplied public HTTPS URLs without a search provider. The reader does not log in, execute page JavaScript, bypass robots rules, or access internal networks. Site restrictions can prevent extraction.
Optional: remote MCP tools
Configure reviewed Streamable HTTP endpoints using a private JSON array. Only the owner workspace can use these shared-credential connectors; guests cannot inherit your external account privileges.
MCP_SERVERS_JSON=[{"id":"my-tools","name":"My approved tools","url":"https://tools.example.com/mcp","token":"private-bearer-token"}]Replace the example with a real compatible service. Discover tool schemas under Connections. Every tool call pauses for exact-argument approval. This client supports public HTTPS and bearer authentication, not an OAuth wizard or local stdio. Connector compatibility must be tested with your chosen server. Remote tools can have external effects; approve only the intended action.
Optional: local Python
Only the local runtime can run approved Python snippets. Install Docker, obtain and review the selected image, then pull it yourself:
docker pull python:3.12-alpine
Add STARLO_DOCKER=1 to .env.local and restart. The container runs without network or host mounts, with a read-only root, unprivileged user, 128 MB memory and a 12-second timeout. The operator may use STARLO_PYTHON_IMAGE to pin a reviewed image digest. This is defense in depth, not a claim of a certified sandbox. Docker is not invoked on Vercel.
Scheduled and background work
The local process advances the owner’s queued missions while it is running, even with the browser closed. Finite schedules specify an interval, a per-run budget, and a maximum execution count. Approval requests stop the relevant mission.
On Vercel, the open browser advances work. For owner work to continue without it, configure an external dispatcher to call /api/tick with Authorization: Bearer YOUR_CRON_SECRET. Set a secret of at least 32 characters. Each invocation advances one bounded action; it is not a continuously running worker. No cron service is automatically provisioned or activated by this ZIP. Consult docs/SCHEDULING.md for the opt-in Vercel configuration and cost implications.
Guest missions require their active browser in this release. Do not promise 24/7 execution unless the appropriate worker is configured and monitored.
Before public commercial use
Complete real provider, database, connector, and target-platform tests. Configure spend alerts, monitoring, backups and incident response. Replace the private-pilot legal pages with reviewed terms and an identified operator/contact. Use a hosting plan that permits your intended business use. Token checkout and signed native releases are not included. The download center carries an unsigned Windows installer candidate with its native-test limitations.
Official setup references
Vercel Drop · Upstash integration · OpenAI function calling · OpenRouter tools · Docker runtime controls