Download. Run.
Inspect the result.
Three useful local workflows work without an API account. Add a model only when you need open-ended language generation and tool decisions.
1. Start the application.
The Windows x64 setup candidate embeds StarloOS and downloads a checksum-pinned official Node runtime on first setup. It is unsigned and not yet executed on Windows in our build environment. Do not disable operating-system protections.
For source users with Node.js 22.16+ in the 22.x line, run npm start. On macOS/Linux without Node, the optional Install-and-Start.sh bootstrap asks before downloading a private runtime. This download path has not been externally exercised.
2. Use the local workbench.
The private workspace and owner access are created automatically. No Redis, Vercel, API key, wallet, or manual administrator setup is required. The browser opens the local /workbench page.
Analyze a table: upload or paste CSV, optionally select a group and numeric measure, and receive calculated statistics and clean CSV exports.
Compare evidence: supply before/after text and receive an exact line-change report, diff, and original input hashes.
Route documents: train an agent first, then classify supplied text using the learned model. Weak matches abstain and ask for human review.
3. Train a narrow local model.
Open Train an agent. Add labelled text examples, with at least three examples per label and at least two labels. Keep evaluation examples in the separate holdout field. Training learns TF-IDF class centroids; evaluation measures performance on those supplied holdout examples.
No model is trained on your holdout examples. Exact duplicate example leakage is rejected. Scores are cosine similarities, not probabilities. This is genuine narrow machine learning, not LLM fine-tuning or proof of broad understanding. You can export and import the examples to retrain elsewhere.
4. Review and download.
Every built-in job produces real files. A completed computation waits for review: accept or reject it, then download individual files or the complete bundle. The receipt records hashes of inputs, outputs, events, and review. It is not a publisher signature or independent proof of correctness.
node scripts/verify-job.mjs /path/to/extracted-mission
Jobs, models, and reviews survive a process restart. The local worker advances jobs only while it runs. Unfinished deterministic work resumes after restart.
5. Connect a model when you need one.
Open Model connection. Configure your OpenAI/OpenRouter account or a separately installed, tool-compatible Ollama model. Keep secrets in memory or explicitly save them to the passphrase-encrypted local vault. Testing sends one small provider request only after you choose it.
Use Advanced command center for model-driven research, writing, tool calls, specialist delegation, approvals, and memory. There is no bundled generative model or prepaid usage. A real funded provider task has not been verified in this environment.
Limits, privacy, and shutdown
Built-in workloads are bounded: CSV up to 50 KB / 1,000 rows; comparisons up to 600 lines per document; 16 trained routers; 50 retained jobs. Export and remove old runs when capacity is reached. Your input stays in the local process for built-in jobs. Optional model requests send their supplied task context to the chosen provider.
Closing the tab leaves the worker running. Finish/cancel outstanding jobs and use Model connection → Stop runtime. Windows data stays separately in %LOCALAPPDATA%\StarloOS\data. There is no automatic updater or native uninstaller. Back up your data before removal. See the package's docs/WORKBENCH.md.
STARLO payments are not enabled. This is an agent application—not a bootable operating system, trading bot, custody service, or promise of returns.