--- title: "OrbySystems OS" url: https://orbysystems.ai/operating-system --- [← Back to the big picture](https://orbysystems.ai/#why) OrbySystems OS # The operating system for _private AI._ Serious AI. On the laptop you already own. OrbySystems brings local models, your documents and useful tools together. A purpose-built operating system orchestrates performance, protects your data and validates the numbers behind the answer. - Works with Wi-Fi off after setup - No metered tokens for local AI - Your documents. Your control. Runs as an application on your existing OS. Personal finance is our first application. The platform reaches further. [Explore the engineering ↘](https://orbysystems.ai/operating-system#systems) [Meet the creators ↘](https://orbysystems.ai/operating-system#creators) OrbySystems Local AI Private by choice. Useful by engineering. 01 / Your system ## One system. AI optimized for your hardware and your data. Layered memory, prompt caching and context management keep useful work ready. Your documents, projects and questions build on each other, so the system carries that work forward. A scheduler and execution pipeline tuned to the compute and memory available put your hardware to work. Less starting over. More moving forward. The model and all its dependencies run in a secure sandbox, and the OS controls access to tools. Its statement agent uses local AI to understand page layouts; OrbySystems handles the download steps. You sign in yourself and can stop the run. **About 97% of prompt tokens reused.** Measured across our laptop and desktop benchmark runs. **Mac** Apple silicon M1 or later, from 8 GB unified memory. **Windows / Linux** Intel or AMD CPUs, with Intel, AMD or NVIDIA GPUs with at least 8 GB VRAM. [Open the OrbySystems Metrics capture](https://orbysystems.ai/launch/assets/orbysystems-pulse-full.jpg) **Pulse & Privacy.**Captured in the app · Synthetic household · Latest Mac · Enlarge ↗ ### Serious local AI. _No per-token bill._ Recorded OrbySystems Medium runs, rounded for orientation. Hardware and workload affect results. Details below. #### Apple silicon Measured M4 Pro → M5 Max Output generation 30–85 tokens/sec Prompt cache hits 97–98% Fresh-input prefill 250–1.1K tokens/sec Measurements & the full performance picture The Apple ranges summarize six archived Medium-model sweeps on M4 Pro (24 GB) and M5 Max (64 GB). Output generation spans per-run medians of 32.4–81.0 tokens/sec; fresh-input prefill spans 266–1,052 tokens/sec; token-weighted prompt reuse spans 97.16–97.63%. The displayed ranges round outward. They are typical recorded rates, not arithmetic averages or measurements of every supported Apple device. The first-text test asks How did my portfolio do this year? against invented household investment data. Each state runs three times; the table shows medians. Cold includes loading the model and preparing its prompt. Warm excludes the preparation already completed in the background: a median 37.5 seconds on M4 Pro and 8.47 seconds on M5 Max. First text is the start of a response, including thinking. It is not a finished answer. | Recorded state | M4 Pro first text / full answer | M5 Max first text / full answer | | --- | --- | --- | | Cold | 41.4 s / 99.1 s | 13.2 s / 23.9 s | | Warm | 1.5 s / 64.8 s | 1.0 s / 11.4 s | | After another chat | 2.5 s / 26.3 s | 0.5 s / 26.3 s | | Saved context restored | 1.0 s / 71.5 s | 1.0 s / 11.4 s | Tool calls and answer lengths vary between states, so first-text improvements do not imply the same improvement in complete-answer time. The timing probe does not independently grade the answer. Cache reuse and prefill come from wider synthetic spending and investment sweeps, with three repeats per task. Prefill excludes cached tokens; generation measures newly produced output. The sweeps include failed attempts: 110 of 141 passed on M4 Pro and 109 of 144 on M5 Max, using different suite versions. These rates are not a claim about every financial answer. The machines were on mains power with no thermal warning reported at the start or end. Results reflect the application, inference runtime, model and hardware together. They do not isolate the contribution of any one optimization or establish server-GPU equivalence. Recorded desktop prefill exceeded 1,000 new input tokens/sec: a 1,029 tokens/sec median on M5 Max, versus 270 on the M4 Pro laptop. Generation medians were 80 and 32 output tokens/sec respectively. The product image is a real capture from a separate local demonstration on 9 November, using the built-in synthetic investment household. It shows a five-call moving average. The 97% headline and comparison table come from the larger archived sweeps described above. ### Your choice of AI. Use the included local AI, connect a compatible open-weight model, or choose a supported cloud provider. Cloud use sends relevant context to that provider; its terms and pricing apply. ### A platform. Finance first. OrbySystems OS is a general-purpose operating system for private AI. Personal finance is our first application, connecting spending, investments, retirement, assets and trusts. General projects bring AI to personal documents on any subject—included in Everyday · Free. Account optional. The creators ### Built by systems engineers. Founder-funded development. The team’s experience spans operating systems, distributed networking, edge computing and cloud infrastructure. [Our story ↗](https://orbysystems.ai/about) [Why Everyday is free ↗](https://orbysystems.ai/plans#founder) 02 / Your data ## Your data. _Protected. Put to work._ Private records become a useful financial picture. Specialist parsers turn PDFs, CSVs and spreadsheet exports into usable records, preserving balances, holdings, income and activity. Recipes, graphs and categories help you slice and dice your data. Drill into a total or chart to see the records behind it. **Your statements. Two ways in.** Add files you’ve downloaded, or let the Statement Download Agent (beta) collect them after you sign in yourself. For an unfamiliar page, show it one download. It learns on your laptop and remembers for next time. The local model interprets the page without receiving your password or clicking through your account. OrbySystems follows the download steps you demonstrated and checks files against the listed account and period. If the page changes, it asks for your help. You can stop at any time. Your choices when you connect online Local analysis works with Wi-Fi off after setup. Downloads, updates, paid activation and enabled reporting connect online. Turning on an online tool gives it a defined connection; choosing a cloud model sends prompts and relevant context to that provider. Privacy and network indicators show the active configuration. [Open the spending overview app capture](https://orbysystems.ai/launch/assets/orbysystems-spending-overview-detail.svg) **Start here → Overview.**Spending & categories · Synthetic household · Enlarge ↗ ### Real formats. Better coverage. Community trust has helped OrbySystems build a large, finance-specific collection of contributed statement formats. Accounts are optional for Everyday. Contributing is always optional: personal details and financial values are replaced on your laptop, and you review the sample before sharing. That real-world diversity powers synthetic test datasets, broader institution coverage and repeatable checks on financial accuracy. [How documents become answers ↗](https://orbysystems.ai/how-it-works#walkthrough) ### Your records stay portable. OrbySystems becomes your local home for financial records, official statements and the history you build over time. Published data contracts, open-source parsers where possible and a community library keep those records understandable and portable. Your financial history belongs to you, wherever you go next. [Community library ↗](https://github.com/OrbySystems/OrbySystems-Finance/tree/main/parsers) [Data formats ↗](https://github.com/OrbySystems/OrbySystems-Finance/blob/main/parsers/IMPORT-CSV.md) 03 / Numbers you can trust ## Numbers you can trust. The answer carries the evidence with it. The OS checks financial inputs and results against source records and calculations. It traces figures through the analysis and marks the answer to show the support behind it. Follow a number back to your statements. Inspect the calculation. See the assumptions behind a scenario. Figures that need review are visibly marked. Traced to evidence Assumptions & estimates Untraced figures [Open the Orby answer with evidence marks](https://orbysystems.ai/launch/assets/orbysystems-evidence-answer-detail.svg) **Ask Orby. Follow the evidence.**Fresh chat · Synthetic household · Enlarge ↗ ## Put private AI to work. Try personal finance, or talk with us about the platform. [Get Everyday · Free](https://orbysystems.ai/get-started) [Talk to the creators ↗](mailto:support@orbysystems.ai?subject=OrbySystems%20platform%20enquiry)